5 * "A fair jaw-cracker dwarf-language must be." --Samwise Gamgee
8 /* This file contains functions for compiling a regular expression. See
9 * also regexec.c which funnily enough, contains functions for executing
10 * a regular expression.
12 * This file is also copied at build time to ext/re/re_comp.c, where
13 * it's built with -DPERL_EXT_RE_BUILD -DPERL_EXT_RE_DEBUG -DPERL_EXT.
14 * This causes the main functions to be compiled under new names and with
15 * debugging support added, which makes "use re 'debug'" work.
18 /* NOTE: this is derived from Henry Spencer's regexp code, and should not
19 * confused with the original package (see point 3 below). Thanks, Henry!
22 /* Additional note: this code is very heavily munged from Henry's version
23 * in places. In some spots I've traded clarity for efficiency, so don't
24 * blame Henry for some of the lack of readability.
27 /* The names of the functions have been changed from regcomp and
28 * regexec to pregcomp and pregexec in order to avoid conflicts
29 * with the POSIX routines of the same names.
32 #ifdef PERL_EXT_RE_BUILD
37 * pregcomp and pregexec -- regsub and regerror are not used in perl
39 * Copyright (c) 1986 by University of Toronto.
40 * Written by Henry Spencer. Not derived from licensed software.
42 * Permission is granted to anyone to use this software for any
43 * purpose on any computer system, and to redistribute it freely,
44 * subject to the following restrictions:
46 * 1. The author is not responsible for the consequences of use of
47 * this software, no matter how awful, even if they arise
50 * 2. The origin of this software must not be misrepresented, either
51 * by explicit claim or by omission.
53 * 3. Altered versions must be plainly marked as such, and must not
54 * be misrepresented as being the original software.
57 **** Alterations to Henry's code are...
59 **** Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
60 **** 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007 by Larry Wall and others
62 **** You may distribute under the terms of either the GNU General Public
63 **** License or the Artistic License, as specified in the README file.
66 * Beware that some of this code is subtly aware of the way operator
67 * precedence is structured in regular expressions. Serious changes in
68 * regular-expression syntax might require a total rethink.
71 #define PERL_IN_REGCOMP_C
74 #ifndef PERL_IN_XSUB_RE
79 #ifdef PERL_IN_XSUB_RE
90 # if defined(BUGGY_MSC6)
91 /* MSC 6.00A breaks on op/regexp.t test 85 unless we turn this off */
92 # pragma optimize("a",off)
93 /* But MSC 6.00A is happy with 'w', for aliases only across function calls*/
94 # pragma optimize("w",on )
95 # endif /* BUGGY_MSC6 */
102 typedef struct RExC_state_t {
103 U32 flags; /* are we folding, multilining? */
104 char *precomp; /* uncompiled string. */
105 REGEXP *rx_sv; /* The SV that is the regexp. */
106 regexp *rx; /* perl core regexp structure */
107 regexp_internal *rxi; /* internal data for regexp object pprivate field */
108 char *start; /* Start of input for compile */
109 char *end; /* End of input for compile */
110 char *parse; /* Input-scan pointer. */
111 I32 whilem_seen; /* number of WHILEM in this expr */
112 regnode *emit_start; /* Start of emitted-code area */
113 regnode *emit_bound; /* First regnode outside of the allocated space */
114 regnode *emit; /* Code-emit pointer; ®dummy = don't = compiling */
115 I32 naughty; /* How bad is this pattern? */
116 I32 sawback; /* Did we see \1, ...? */
118 I32 size; /* Code size. */
119 I32 npar; /* Capture buffer count, (OPEN). */
120 I32 cpar; /* Capture buffer count, (CLOSE). */
121 I32 nestroot; /* root parens we are in - used by accept */
125 regnode **open_parens; /* pointers to open parens */
126 regnode **close_parens; /* pointers to close parens */
127 regnode *opend; /* END node in program */
128 I32 utf8; /* whether the pattern is utf8 or not */
129 I32 orig_utf8; /* whether the pattern was originally in utf8 */
130 /* XXX use this for future optimisation of case
131 * where pattern must be upgraded to utf8. */
132 HV *charnames; /* cache of named sequences */
133 HV *paren_names; /* Paren names */
135 regnode **recurse; /* Recurse regops */
136 I32 recurse_count; /* Number of recurse regops */
138 char *starttry; /* -Dr: where regtry was called. */
139 #define RExC_starttry (pRExC_state->starttry)
142 const char *lastparse;
144 AV *paren_name_list; /* idx -> name */
145 #define RExC_lastparse (pRExC_state->lastparse)
146 #define RExC_lastnum (pRExC_state->lastnum)
147 #define RExC_paren_name_list (pRExC_state->paren_name_list)
151 #define RExC_flags (pRExC_state->flags)
152 #define RExC_precomp (pRExC_state->precomp)
153 #define RExC_rx_sv (pRExC_state->rx_sv)
154 #define RExC_rx (pRExC_state->rx)
155 #define RExC_rxi (pRExC_state->rxi)
156 #define RExC_start (pRExC_state->start)
157 #define RExC_end (pRExC_state->end)
158 #define RExC_parse (pRExC_state->parse)
159 #define RExC_whilem_seen (pRExC_state->whilem_seen)
160 #ifdef RE_TRACK_PATTERN_OFFSETS
161 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
163 #define RExC_emit (pRExC_state->emit)
164 #define RExC_emit_start (pRExC_state->emit_start)
165 #define RExC_emit_bound (pRExC_state->emit_bound)
166 #define RExC_naughty (pRExC_state->naughty)
167 #define RExC_sawback (pRExC_state->sawback)
168 #define RExC_seen (pRExC_state->seen)
169 #define RExC_size (pRExC_state->size)
170 #define RExC_npar (pRExC_state->npar)
171 #define RExC_nestroot (pRExC_state->nestroot)
172 #define RExC_extralen (pRExC_state->extralen)
173 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
174 #define RExC_seen_evals (pRExC_state->seen_evals)
175 #define RExC_utf8 (pRExC_state->utf8)
176 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
177 #define RExC_charnames (pRExC_state->charnames)
178 #define RExC_open_parens (pRExC_state->open_parens)
179 #define RExC_close_parens (pRExC_state->close_parens)
180 #define RExC_opend (pRExC_state->opend)
181 #define RExC_paren_names (pRExC_state->paren_names)
182 #define RExC_recurse (pRExC_state->recurse)
183 #define RExC_recurse_count (pRExC_state->recurse_count)
186 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
187 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
188 ((*s) == '{' && regcurly(s)))
191 #undef SPSTART /* dratted cpp namespace... */
194 * Flags to be passed up and down.
196 #define WORST 0 /* Worst case. */
197 #define HASWIDTH 0x01 /* Known to match non-null strings. */
198 #define SIMPLE 0x02 /* Simple enough to be STAR/PLUS operand. */
199 #define SPSTART 0x04 /* Starts with * or +. */
200 #define TRYAGAIN 0x08 /* Weeded out a declaration. */
201 #define POSTPONED 0x10 /* (?1),(?&name), (??{...}) or similar */
203 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
205 /* whether trie related optimizations are enabled */
206 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
207 #define TRIE_STUDY_OPT
208 #define FULL_TRIE_STUDY
214 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
215 #define PBITVAL(paren) (1 << ((paren) & 7))
216 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
217 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
218 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
221 /* About scan_data_t.
223 During optimisation we recurse through the regexp program performing
224 various inplace (keyhole style) optimisations. In addition study_chunk
225 and scan_commit populate this data structure with information about
226 what strings MUST appear in the pattern. We look for the longest
227 string that must appear for at a fixed location, and we look for the
228 longest string that may appear at a floating location. So for instance
233 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
234 strings (because they follow a .* construct). study_chunk will identify
235 both FOO and BAR as being the longest fixed and floating strings respectively.
237 The strings can be composites, for instance
241 will result in a composite fixed substring 'foo'.
243 For each string some basic information is maintained:
245 - offset or min_offset
246 This is the position the string must appear at, or not before.
247 It also implicitly (when combined with minlenp) tells us how many
248 character must match before the string we are searching.
249 Likewise when combined with minlenp and the length of the string
250 tells us how many characters must appear after the string we have
254 Only used for floating strings. This is the rightmost point that
255 the string can appear at. Ifset to I32 max it indicates that the
256 string can occur infinitely far to the right.
259 A pointer to the minimum length of the pattern that the string
260 was found inside. This is important as in the case of positive
261 lookahead or positive lookbehind we can have multiple patterns
266 The minimum length of the pattern overall is 3, the minimum length
267 of the lookahead part is 3, but the minimum length of the part that
268 will actually match is 1. So 'FOO's minimum length is 3, but the
269 minimum length for the F is 1. This is important as the minimum length
270 is used to determine offsets in front of and behind the string being
271 looked for. Since strings can be composites this is the length of the
272 pattern at the time it was commited with a scan_commit. Note that
273 the length is calculated by study_chunk, so that the minimum lengths
274 are not known until the full pattern has been compiled, thus the
275 pointer to the value.
279 In the case of lookbehind the string being searched for can be
280 offset past the start point of the final matching string.
281 If this value was just blithely removed from the min_offset it would
282 invalidate some of the calculations for how many chars must match
283 before or after (as they are derived from min_offset and minlen and
284 the length of the string being searched for).
285 When the final pattern is compiled and the data is moved from the
286 scan_data_t structure into the regexp structure the information
287 about lookbehind is factored in, with the information that would
288 have been lost precalculated in the end_shift field for the
291 The fields pos_min and pos_delta are used to store the minimum offset
292 and the delta to the maximum offset at the current point in the pattern.
296 typedef struct scan_data_t {
297 /*I32 len_min; unused */
298 /*I32 len_delta; unused */
302 I32 last_end; /* min value, <0 unless valid. */
305 SV **longest; /* Either &l_fixed, or &l_float. */
306 SV *longest_fixed; /* longest fixed string found in pattern */
307 I32 offset_fixed; /* offset where it starts */
308 I32 *minlen_fixed; /* pointer to the minlen relevent to the string */
309 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
310 SV *longest_float; /* longest floating string found in pattern */
311 I32 offset_float_min; /* earliest point in string it can appear */
312 I32 offset_float_max; /* latest point in string it can appear */
313 I32 *minlen_float; /* pointer to the minlen relevent to the string */
314 I32 lookbehind_float; /* is the position of the string modified by LB */
318 struct regnode_charclass_class *start_class;
322 * Forward declarations for pregcomp()'s friends.
325 static const scan_data_t zero_scan_data =
326 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
328 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
329 #define SF_BEFORE_SEOL 0x0001
330 #define SF_BEFORE_MEOL 0x0002
331 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
332 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
335 # define SF_FIX_SHIFT_EOL (0+2)
336 # define SF_FL_SHIFT_EOL (0+4)
338 # define SF_FIX_SHIFT_EOL (+2)
339 # define SF_FL_SHIFT_EOL (+4)
342 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
343 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
345 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
346 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
347 #define SF_IS_INF 0x0040
348 #define SF_HAS_PAR 0x0080
349 #define SF_IN_PAR 0x0100
350 #define SF_HAS_EVAL 0x0200
351 #define SCF_DO_SUBSTR 0x0400
352 #define SCF_DO_STCLASS_AND 0x0800
353 #define SCF_DO_STCLASS_OR 0x1000
354 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
355 #define SCF_WHILEM_VISITED_POS 0x2000
357 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
358 #define SCF_SEEN_ACCEPT 0x8000
360 #define UTF (RExC_utf8 != 0)
361 #define LOC ((RExC_flags & RXf_PMf_LOCALE) != 0)
362 #define FOLD ((RExC_flags & RXf_PMf_FOLD) != 0)
364 #define OOB_UNICODE 12345678
365 #define OOB_NAMEDCLASS -1
367 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
368 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
371 /* length of regex to show in messages that don't mark a position within */
372 #define RegexLengthToShowInErrorMessages 127
375 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
376 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
377 * op/pragma/warn/regcomp.
379 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
380 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
382 #define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
385 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
386 * arg. Show regex, up to a maximum length. If it's too long, chop and add
389 #define _FAIL(code) STMT_START { \
390 const char *ellipses = ""; \
391 IV len = RExC_end - RExC_precomp; \
394 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
395 if (len > RegexLengthToShowInErrorMessages) { \
396 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
397 len = RegexLengthToShowInErrorMessages - 10; \
403 #define FAIL(msg) _FAIL( \
404 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
405 msg, (int)len, RExC_precomp, ellipses))
407 #define FAIL2(msg,arg) _FAIL( \
408 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
409 arg, (int)len, RExC_precomp, ellipses))
412 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
414 #define Simple_vFAIL(m) STMT_START { \
415 const IV offset = RExC_parse - RExC_precomp; \
416 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
417 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
421 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
423 #define vFAIL(m) STMT_START { \
425 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
430 * Like Simple_vFAIL(), but accepts two arguments.
432 #define Simple_vFAIL2(m,a1) STMT_START { \
433 const IV offset = RExC_parse - RExC_precomp; \
434 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
435 (int)offset, RExC_precomp, RExC_precomp + offset); \
439 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
441 #define vFAIL2(m,a1) STMT_START { \
443 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
444 Simple_vFAIL2(m, a1); \
449 * Like Simple_vFAIL(), but accepts three arguments.
451 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
452 const IV offset = RExC_parse - RExC_precomp; \
453 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
454 (int)offset, RExC_precomp, RExC_precomp + offset); \
458 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
460 #define vFAIL3(m,a1,a2) STMT_START { \
462 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
463 Simple_vFAIL3(m, a1, a2); \
467 * Like Simple_vFAIL(), but accepts four arguments.
469 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
470 const IV offset = RExC_parse - RExC_precomp; \
471 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
472 (int)offset, RExC_precomp, RExC_precomp + offset); \
475 #define vWARN(loc,m) STMT_START { \
476 const IV offset = loc - RExC_precomp; \
477 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s" REPORT_LOCATION, \
478 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
481 #define vWARNdep(loc,m) STMT_START { \
482 const IV offset = loc - RExC_precomp; \
483 Perl_warner(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
484 "%s" REPORT_LOCATION, \
485 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
489 #define vWARN2(loc, m, a1) STMT_START { \
490 const IV offset = loc - RExC_precomp; \
491 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
492 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
495 #define vWARN3(loc, m, a1, a2) STMT_START { \
496 const IV offset = loc - RExC_precomp; \
497 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
498 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
501 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
502 const IV offset = loc - RExC_precomp; \
503 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
504 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
507 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
508 const IV offset = loc - RExC_precomp; \
509 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
510 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
514 /* Allow for side effects in s */
515 #define REGC(c,s) STMT_START { \
516 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
519 /* Macros for recording node offsets. 20001227 mjd@plover.com
520 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
521 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
522 * Element 0 holds the number n.
523 * Position is 1 indexed.
525 #ifndef RE_TRACK_PATTERN_OFFSETS
526 #define Set_Node_Offset_To_R(node,byte)
527 #define Set_Node_Offset(node,byte)
528 #define Set_Cur_Node_Offset
529 #define Set_Node_Length_To_R(node,len)
530 #define Set_Node_Length(node,len)
531 #define Set_Node_Cur_Length(node)
532 #define Node_Offset(n)
533 #define Node_Length(n)
534 #define Set_Node_Offset_Length(node,offset,len)
535 #define ProgLen(ri) ri->u.proglen
536 #define SetProgLen(ri,x) ri->u.proglen = x
538 #define ProgLen(ri) ri->u.offsets[0]
539 #define SetProgLen(ri,x) ri->u.offsets[0] = x
540 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
542 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
543 __LINE__, (int)(node), (int)(byte))); \
545 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
547 RExC_offsets[2*(node)-1] = (byte); \
552 #define Set_Node_Offset(node,byte) \
553 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
554 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
556 #define Set_Node_Length_To_R(node,len) STMT_START { \
558 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
559 __LINE__, (int)(node), (int)(len))); \
561 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
563 RExC_offsets[2*(node)] = (len); \
568 #define Set_Node_Length(node,len) \
569 Set_Node_Length_To_R((node)-RExC_emit_start, len)
570 #define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
571 #define Set_Node_Cur_Length(node) \
572 Set_Node_Length(node, RExC_parse - parse_start)
574 /* Get offsets and lengths */
575 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
576 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
578 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
579 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
580 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
584 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
585 #define EXPERIMENTAL_INPLACESCAN
586 #endif /*RE_TRACK_PATTERN_OFFSETS*/
588 #define DEBUG_STUDYDATA(str,data,depth) \
589 DEBUG_OPTIMISE_MORE_r(if(data){ \
590 PerlIO_printf(Perl_debug_log, \
591 "%*s" str "Pos:%"IVdf"/%"IVdf \
592 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
593 (int)(depth)*2, "", \
594 (IV)((data)->pos_min), \
595 (IV)((data)->pos_delta), \
596 (UV)((data)->flags), \
597 (IV)((data)->whilem_c), \
598 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
599 is_inf ? "INF " : "" \
601 if ((data)->last_found) \
602 PerlIO_printf(Perl_debug_log, \
603 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
604 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
605 SvPVX_const((data)->last_found), \
606 (IV)((data)->last_end), \
607 (IV)((data)->last_start_min), \
608 (IV)((data)->last_start_max), \
609 ((data)->longest && \
610 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
611 SvPVX_const((data)->longest_fixed), \
612 (IV)((data)->offset_fixed), \
613 ((data)->longest && \
614 (data)->longest==&((data)->longest_float)) ? "*" : "", \
615 SvPVX_const((data)->longest_float), \
616 (IV)((data)->offset_float_min), \
617 (IV)((data)->offset_float_max) \
619 PerlIO_printf(Perl_debug_log,"\n"); \
622 static void clear_re(pTHX_ void *r);
624 /* Mark that we cannot extend a found fixed substring at this point.
625 Update the longest found anchored substring and the longest found
626 floating substrings if needed. */
629 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
631 const STRLEN l = CHR_SVLEN(data->last_found);
632 const STRLEN old_l = CHR_SVLEN(*data->longest);
633 GET_RE_DEBUG_FLAGS_DECL;
635 PERL_ARGS_ASSERT_SCAN_COMMIT;
637 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
638 SvSetMagicSV(*data->longest, data->last_found);
639 if (*data->longest == data->longest_fixed) {
640 data->offset_fixed = l ? data->last_start_min : data->pos_min;
641 if (data->flags & SF_BEFORE_EOL)
643 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
645 data->flags &= ~SF_FIX_BEFORE_EOL;
646 data->minlen_fixed=minlenp;
647 data->lookbehind_fixed=0;
649 else { /* *data->longest == data->longest_float */
650 data->offset_float_min = l ? data->last_start_min : data->pos_min;
651 data->offset_float_max = (l
652 ? data->last_start_max
653 : data->pos_min + data->pos_delta);
654 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
655 data->offset_float_max = I32_MAX;
656 if (data->flags & SF_BEFORE_EOL)
658 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
660 data->flags &= ~SF_FL_BEFORE_EOL;
661 data->minlen_float=minlenp;
662 data->lookbehind_float=0;
665 SvCUR_set(data->last_found, 0);
667 SV * const sv = data->last_found;
668 if (SvUTF8(sv) && SvMAGICAL(sv)) {
669 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
675 data->flags &= ~SF_BEFORE_EOL;
676 DEBUG_STUDYDATA("commit: ",data,0);
679 /* Can match anything (initialization) */
681 S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
683 PERL_ARGS_ASSERT_CL_ANYTHING;
685 ANYOF_CLASS_ZERO(cl);
686 ANYOF_BITMAP_SETALL(cl);
687 cl->flags = ANYOF_EOS|ANYOF_UNICODE_ALL;
689 cl->flags |= ANYOF_LOCALE;
692 /* Can match anything (initialization) */
694 S_cl_is_anything(const struct regnode_charclass_class *cl)
698 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
700 for (value = 0; value <= ANYOF_MAX; value += 2)
701 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
703 if (!(cl->flags & ANYOF_UNICODE_ALL))
705 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
710 /* Can match anything (initialization) */
712 S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
714 PERL_ARGS_ASSERT_CL_INIT;
716 Zero(cl, 1, struct regnode_charclass_class);
718 cl_anything(pRExC_state, cl);
722 S_cl_init_zero(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
724 PERL_ARGS_ASSERT_CL_INIT_ZERO;
726 Zero(cl, 1, struct regnode_charclass_class);
728 cl_anything(pRExC_state, cl);
730 cl->flags |= ANYOF_LOCALE;
733 /* 'And' a given class with another one. Can create false positives */
734 /* We assume that cl is not inverted */
736 S_cl_and(struct regnode_charclass_class *cl,
737 const struct regnode_charclass_class *and_with)
739 PERL_ARGS_ASSERT_CL_AND;
741 assert(and_with->type == ANYOF);
742 if (!(and_with->flags & ANYOF_CLASS)
743 && !(cl->flags & ANYOF_CLASS)
744 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
745 && !(and_with->flags & ANYOF_FOLD)
746 && !(cl->flags & ANYOF_FOLD)) {
749 if (and_with->flags & ANYOF_INVERT)
750 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
751 cl->bitmap[i] &= ~and_with->bitmap[i];
753 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
754 cl->bitmap[i] &= and_with->bitmap[i];
755 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
756 if (!(and_with->flags & ANYOF_EOS))
757 cl->flags &= ~ANYOF_EOS;
759 if (cl->flags & ANYOF_UNICODE_ALL && and_with->flags & ANYOF_UNICODE &&
760 !(and_with->flags & ANYOF_INVERT)) {
761 cl->flags &= ~ANYOF_UNICODE_ALL;
762 cl->flags |= ANYOF_UNICODE;
763 ARG_SET(cl, ARG(and_with));
765 if (!(and_with->flags & ANYOF_UNICODE_ALL) &&
766 !(and_with->flags & ANYOF_INVERT))
767 cl->flags &= ~ANYOF_UNICODE_ALL;
768 if (!(and_with->flags & (ANYOF_UNICODE|ANYOF_UNICODE_ALL)) &&
769 !(and_with->flags & ANYOF_INVERT))
770 cl->flags &= ~ANYOF_UNICODE;
773 /* 'OR' a given class with another one. Can create false positives */
774 /* We assume that cl is not inverted */
776 S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
778 PERL_ARGS_ASSERT_CL_OR;
780 if (or_with->flags & ANYOF_INVERT) {
782 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
783 * <= (B1 | !B2) | (CL1 | !CL2)
784 * which is wasteful if CL2 is small, but we ignore CL2:
785 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
786 * XXXX Can we handle case-fold? Unclear:
787 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
788 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
790 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
791 && !(or_with->flags & ANYOF_FOLD)
792 && !(cl->flags & ANYOF_FOLD) ) {
795 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
796 cl->bitmap[i] |= ~or_with->bitmap[i];
797 } /* XXXX: logic is complicated otherwise */
799 cl_anything(pRExC_state, cl);
802 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
803 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
804 && (!(or_with->flags & ANYOF_FOLD)
805 || (cl->flags & ANYOF_FOLD)) ) {
808 /* OR char bitmap and class bitmap separately */
809 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
810 cl->bitmap[i] |= or_with->bitmap[i];
811 if (or_with->flags & ANYOF_CLASS) {
812 for (i = 0; i < ANYOF_CLASSBITMAP_SIZE; i++)
813 cl->classflags[i] |= or_with->classflags[i];
814 cl->flags |= ANYOF_CLASS;
817 else { /* XXXX: logic is complicated, leave it along for a moment. */
818 cl_anything(pRExC_state, cl);
821 if (or_with->flags & ANYOF_EOS)
822 cl->flags |= ANYOF_EOS;
824 if (cl->flags & ANYOF_UNICODE && or_with->flags & ANYOF_UNICODE &&
825 ARG(cl) != ARG(or_with)) {
826 cl->flags |= ANYOF_UNICODE_ALL;
827 cl->flags &= ~ANYOF_UNICODE;
829 if (or_with->flags & ANYOF_UNICODE_ALL) {
830 cl->flags |= ANYOF_UNICODE_ALL;
831 cl->flags &= ~ANYOF_UNICODE;
835 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
836 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
837 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
838 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
843 dump_trie(trie,widecharmap,revcharmap)
844 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
845 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
847 These routines dump out a trie in a somewhat readable format.
848 The _interim_ variants are used for debugging the interim
849 tables that are used to generate the final compressed
850 representation which is what dump_trie expects.
852 Part of the reason for their existance is to provide a form
853 of documentation as to how the different representations function.
858 Dumps the final compressed table form of the trie to Perl_debug_log.
859 Used for debugging make_trie().
863 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
864 AV *revcharmap, U32 depth)
867 SV *sv=sv_newmortal();
868 int colwidth= widecharmap ? 6 : 4;
869 GET_RE_DEBUG_FLAGS_DECL;
871 PERL_ARGS_ASSERT_DUMP_TRIE;
873 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
874 (int)depth * 2 + 2,"",
875 "Match","Base","Ofs" );
877 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
878 SV ** const tmp = av_fetch( revcharmap, state, 0);
880 PerlIO_printf( Perl_debug_log, "%*s",
882 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
883 PL_colors[0], PL_colors[1],
884 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
885 PERL_PV_ESCAPE_FIRSTCHAR
890 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
891 (int)depth * 2 + 2,"");
893 for( state = 0 ; state < trie->uniquecharcount ; state++ )
894 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
895 PerlIO_printf( Perl_debug_log, "\n");
897 for( state = 1 ; state < trie->statecount ; state++ ) {
898 const U32 base = trie->states[ state ].trans.base;
900 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
902 if ( trie->states[ state ].wordnum ) {
903 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
905 PerlIO_printf( Perl_debug_log, "%6s", "" );
908 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
913 while( ( base + ofs < trie->uniquecharcount ) ||
914 ( base + ofs - trie->uniquecharcount < trie->lasttrans
915 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
918 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
920 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
921 if ( ( base + ofs >= trie->uniquecharcount ) &&
922 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
923 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
925 PerlIO_printf( Perl_debug_log, "%*"UVXf,
927 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
929 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
933 PerlIO_printf( Perl_debug_log, "]");
936 PerlIO_printf( Perl_debug_log, "\n" );
940 Dumps a fully constructed but uncompressed trie in list form.
941 List tries normally only are used for construction when the number of
942 possible chars (trie->uniquecharcount) is very high.
943 Used for debugging make_trie().
946 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
947 HV *widecharmap, AV *revcharmap, U32 next_alloc,
951 SV *sv=sv_newmortal();
952 int colwidth= widecharmap ? 6 : 4;
953 GET_RE_DEBUG_FLAGS_DECL;
955 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
957 /* print out the table precompression. */
958 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
959 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
960 "------:-----+-----------------\n" );
962 for( state=1 ; state < next_alloc ; state ++ ) {
965 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
966 (int)depth * 2 + 2,"", (UV)state );
967 if ( ! trie->states[ state ].wordnum ) {
968 PerlIO_printf( Perl_debug_log, "%5s| ","");
970 PerlIO_printf( Perl_debug_log, "W%4x| ",
971 trie->states[ state ].wordnum
974 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
975 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
977 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
979 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
980 PL_colors[0], PL_colors[1],
981 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
982 PERL_PV_ESCAPE_FIRSTCHAR
984 TRIE_LIST_ITEM(state,charid).forid,
985 (UV)TRIE_LIST_ITEM(state,charid).newstate
988 PerlIO_printf(Perl_debug_log, "\n%*s| ",
989 (int)((depth * 2) + 14), "");
992 PerlIO_printf( Perl_debug_log, "\n");
997 Dumps a fully constructed but uncompressed trie in table form.
998 This is the normal DFA style state transition table, with a few
999 twists to facilitate compression later.
1000 Used for debugging make_trie().
1003 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1004 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1009 SV *sv=sv_newmortal();
1010 int colwidth= widecharmap ? 6 : 4;
1011 GET_RE_DEBUG_FLAGS_DECL;
1013 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1016 print out the table precompression so that we can do a visual check
1017 that they are identical.
1020 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1022 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1023 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1025 PerlIO_printf( Perl_debug_log, "%*s",
1027 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1028 PL_colors[0], PL_colors[1],
1029 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1030 PERL_PV_ESCAPE_FIRSTCHAR
1036 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1038 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1039 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1042 PerlIO_printf( Perl_debug_log, "\n" );
1044 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1046 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1047 (int)depth * 2 + 2,"",
1048 (UV)TRIE_NODENUM( state ) );
1050 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1051 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1053 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1055 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1057 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1058 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1060 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1061 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1068 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1069 startbranch: the first branch in the whole branch sequence
1070 first : start branch of sequence of branch-exact nodes.
1071 May be the same as startbranch
1072 last : Thing following the last branch.
1073 May be the same as tail.
1074 tail : item following the branch sequence
1075 count : words in the sequence
1076 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1077 depth : indent depth
1079 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1081 A trie is an N'ary tree where the branches are determined by digital
1082 decomposition of the key. IE, at the root node you look up the 1st character and
1083 follow that branch repeat until you find the end of the branches. Nodes can be
1084 marked as "accepting" meaning they represent a complete word. Eg:
1088 would convert into the following structure. Numbers represent states, letters
1089 following numbers represent valid transitions on the letter from that state, if
1090 the number is in square brackets it represents an accepting state, otherwise it
1091 will be in parenthesis.
1093 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1097 (1) +-i->(6)-+-s->[7]
1099 +-s->(3)-+-h->(4)-+-e->[5]
1101 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1103 This shows that when matching against the string 'hers' we will begin at state 1
1104 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1105 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1106 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1107 single traverse. We store a mapping from accepting to state to which word was
1108 matched, and then when we have multiple possibilities we try to complete the
1109 rest of the regex in the order in which they occured in the alternation.
1111 The only prior NFA like behaviour that would be changed by the TRIE support is
1112 the silent ignoring of duplicate alternations which are of the form:
1114 / (DUPE|DUPE) X? (?{ ... }) Y /x
1116 Thus EVAL blocks follwing a trie may be called a different number of times with
1117 and without the optimisation. With the optimisations dupes will be silently
1118 ignored. This inconsistant behaviour of EVAL type nodes is well established as
1119 the following demonstrates:
1121 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1123 which prints out 'word' three times, but
1125 'words'=~/(word|word|word)(?{ print $1 })S/
1127 which doesnt print it out at all. This is due to other optimisations kicking in.
1129 Example of what happens on a structural level:
1131 The regexp /(ac|ad|ab)+/ will produce the folowing debug output:
1133 1: CURLYM[1] {1,32767}(18)
1144 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1145 and should turn into:
1147 1: CURLYM[1] {1,32767}(18)
1149 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1157 Cases where tail != last would be like /(?foo|bar)baz/:
1167 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1168 and would end up looking like:
1171 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1178 d = uvuni_to_utf8_flags(d, uv, 0);
1180 is the recommended Unicode-aware way of saying
1185 #define TRIE_STORE_REVCHAR \
1188 SV *zlopp = newSV(2); \
1189 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1190 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, uvc & 0xFF); \
1191 SvCUR_set(zlopp, kapow - flrbbbbb); \
1194 av_push(revcharmap, zlopp); \
1196 char ooooff = (char)uvc; \
1197 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1201 #define TRIE_READ_CHAR STMT_START { \
1205 if ( foldlen > 0 ) { \
1206 uvc = utf8n_to_uvuni( scan, UTF8_MAXLEN, &len, uniflags ); \
1211 uvc = utf8n_to_uvuni( (const U8*)uc, UTF8_MAXLEN, &len, uniflags);\
1212 uvc = to_uni_fold( uvc, foldbuf, &foldlen ); \
1213 foldlen -= UNISKIP( uvc ); \
1214 scan = foldbuf + UNISKIP( uvc ); \
1217 uvc = utf8n_to_uvuni( (const U8*)uc, UTF8_MAXLEN, &len, uniflags);\
1227 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1228 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1229 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1230 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1232 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1233 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1234 TRIE_LIST_CUR( state )++; \
1237 #define TRIE_LIST_NEW(state) STMT_START { \
1238 Newxz( trie->states[ state ].trans.list, \
1239 4, reg_trie_trans_le ); \
1240 TRIE_LIST_CUR( state ) = 1; \
1241 TRIE_LIST_LEN( state ) = 4; \
1244 #define TRIE_HANDLE_WORD(state) STMT_START { \
1245 U16 dupe= trie->states[ state ].wordnum; \
1246 regnode * const noper_next = regnext( noper ); \
1248 if (trie->wordlen) \
1249 trie->wordlen[ curword ] = wordlen; \
1251 /* store the word for dumping */ \
1253 if (OP(noper) != NOTHING) \
1254 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1256 tmp = newSVpvn_utf8( "", 0, UTF ); \
1257 av_push( trie_words, tmp ); \
1262 if ( noper_next < tail ) { \
1264 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
1265 trie->jump[curword] = (U16)(noper_next - convert); \
1267 jumper = noper_next; \
1269 nextbranch= regnext(cur); \
1273 /* So it's a dupe. This means we need to maintain a */\
1274 /* linked-list from the first to the next. */\
1275 /* we only allocate the nextword buffer when there */\
1276 /* a dupe, so first time we have to do the allocation */\
1277 if (!trie->nextword) \
1278 trie->nextword = (U16 *) \
1279 PerlMemShared_calloc( word_count + 1, sizeof(U16)); \
1280 while ( trie->nextword[dupe] ) \
1281 dupe= trie->nextword[dupe]; \
1282 trie->nextword[dupe]= curword; \
1284 /* we haven't inserted this word yet. */ \
1285 trie->states[ state ].wordnum = curword; \
1290 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1291 ( ( base + charid >= ucharcount \
1292 && base + charid < ubound \
1293 && state == trie->trans[ base - ucharcount + charid ].check \
1294 && trie->trans[ base - ucharcount + charid ].next ) \
1295 ? trie->trans[ base - ucharcount + charid ].next \
1296 : ( state==1 ? special : 0 ) \
1300 #define MADE_JUMP_TRIE 2
1301 #define MADE_EXACT_TRIE 4
1304 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
1307 /* first pass, loop through and scan words */
1308 reg_trie_data *trie;
1309 HV *widecharmap = NULL;
1310 AV *revcharmap = newAV();
1312 const U32 uniflags = UTF8_ALLOW_DEFAULT;
1317 regnode *jumper = NULL;
1318 regnode *nextbranch = NULL;
1319 regnode *convert = NULL;
1320 /* we just use folder as a flag in utf8 */
1321 const U8 * const folder = ( flags == EXACTF
1323 : ( flags == EXACTFL
1330 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1331 AV *trie_words = NULL;
1332 /* along with revcharmap, this only used during construction but both are
1333 * useful during debugging so we store them in the struct when debugging.
1336 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
1337 STRLEN trie_charcount=0;
1339 SV *re_trie_maxbuff;
1340 GET_RE_DEBUG_FLAGS_DECL;
1342 PERL_ARGS_ASSERT_MAKE_TRIE;
1344 PERL_UNUSED_ARG(depth);
1347 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1349 trie->startstate = 1;
1350 trie->wordcount = word_count;
1351 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1352 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
1353 if (!(UTF && folder))
1354 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
1356 trie_words = newAV();
1359 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
1360 if (!SvIOK(re_trie_maxbuff)) {
1361 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
1364 PerlIO_printf( Perl_debug_log,
1365 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
1366 (int)depth * 2 + 2, "",
1367 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
1368 REG_NODE_NUM(last), REG_NODE_NUM(tail),
1372 /* Find the node we are going to overwrite */
1373 if ( first == startbranch && OP( last ) != BRANCH ) {
1374 /* whole branch chain */
1377 /* branch sub-chain */
1378 convert = NEXTOPER( first );
1381 /* -- First loop and Setup --
1383 We first traverse the branches and scan each word to determine if it
1384 contains widechars, and how many unique chars there are, this is
1385 important as we have to build a table with at least as many columns as we
1388 We use an array of integers to represent the character codes 0..255
1389 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
1390 native representation of the character value as the key and IV's for the
1393 *TODO* If we keep track of how many times each character is used we can
1394 remap the columns so that the table compression later on is more
1395 efficient in terms of memory by ensuring most common value is in the
1396 middle and the least common are on the outside. IMO this would be better
1397 than a most to least common mapping as theres a decent chance the most
1398 common letter will share a node with the least common, meaning the node
1399 will not be compressable. With a middle is most common approach the worst
1400 case is when we have the least common nodes twice.
1404 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1405 regnode * const noper = NEXTOPER( cur );
1406 const U8 *uc = (U8*)STRING( noper );
1407 const U8 * const e = uc + STR_LEN( noper );
1409 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1410 const U8 *scan = (U8*)NULL;
1411 U32 wordlen = 0; /* required init */
1413 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
1415 if (OP(noper) == NOTHING) {
1419 if ( set_bit ) /* bitmap only alloced when !(UTF&&Folding) */
1420 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1421 regardless of encoding */
1423 for ( ; uc < e ; uc += len ) {
1424 TRIE_CHARCOUNT(trie)++;
1428 if ( !trie->charmap[ uvc ] ) {
1429 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1431 trie->charmap[ folder[ uvc ] ] = trie->charmap[ uvc ];
1435 /* store the codepoint in the bitmap, and if its ascii
1436 also store its folded equivelent. */
1437 TRIE_BITMAP_SET(trie,uvc);
1439 /* store the folded codepoint */
1440 if ( folder ) TRIE_BITMAP_SET(trie,folder[ uvc ]);
1443 /* store first byte of utf8 representation of
1444 codepoints in the 127 < uvc < 256 range */
1445 if (127 < uvc && uvc < 192) {
1446 TRIE_BITMAP_SET(trie,194);
1447 } else if (191 < uvc ) {
1448 TRIE_BITMAP_SET(trie,195);
1449 /* && uvc < 256 -- we know uvc is < 256 already */
1452 set_bit = 0; /* We've done our bit :-) */
1457 widecharmap = newHV();
1459 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
1462 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
1464 if ( !SvTRUE( *svpp ) ) {
1465 sv_setiv( *svpp, ++trie->uniquecharcount );
1470 if( cur == first ) {
1473 } else if (chars < trie->minlen) {
1475 } else if (chars > trie->maxlen) {
1479 } /* end first pass */
1480 DEBUG_TRIE_COMPILE_r(
1481 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1482 (int)depth * 2 + 2,"",
1483 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
1484 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1485 (int)trie->minlen, (int)trie->maxlen )
1487 trie->wordlen = (U32 *) PerlMemShared_calloc( word_count, sizeof(U32) );
1490 We now know what we are dealing with in terms of unique chars and
1491 string sizes so we can calculate how much memory a naive
1492 representation using a flat table will take. If it's over a reasonable
1493 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
1494 conservative but potentially much slower representation using an array
1497 At the end we convert both representations into the same compressed
1498 form that will be used in regexec.c for matching with. The latter
1499 is a form that cannot be used to construct with but has memory
1500 properties similar to the list form and access properties similar
1501 to the table form making it both suitable for fast searches and
1502 small enough that its feasable to store for the duration of a program.
1504 See the comment in the code where the compressed table is produced
1505 inplace from the flat tabe representation for an explanation of how
1506 the compression works.
1511 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
1513 Second Pass -- Array Of Lists Representation
1515 Each state will be represented by a list of charid:state records
1516 (reg_trie_trans_le) the first such element holds the CUR and LEN
1517 points of the allocated array. (See defines above).
1519 We build the initial structure using the lists, and then convert
1520 it into the compressed table form which allows faster lookups
1521 (but cant be modified once converted).
1524 STRLEN transcount = 1;
1526 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1527 "%*sCompiling trie using list compiler\n",
1528 (int)depth * 2 + 2, ""));
1530 trie->states = (reg_trie_state *)
1531 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1532 sizeof(reg_trie_state) );
1536 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1538 regnode * const noper = NEXTOPER( cur );
1539 U8 *uc = (U8*)STRING( noper );
1540 const U8 * const e = uc + STR_LEN( noper );
1541 U32 state = 1; /* required init */
1542 U16 charid = 0; /* sanity init */
1543 U8 *scan = (U8*)NULL; /* sanity init */
1544 STRLEN foldlen = 0; /* required init */
1545 U32 wordlen = 0; /* required init */
1546 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1548 if (OP(noper) != NOTHING) {
1549 for ( ; uc < e ; uc += len ) {
1554 charid = trie->charmap[ uvc ];
1556 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1560 charid=(U16)SvIV( *svpp );
1563 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1570 if ( !trie->states[ state ].trans.list ) {
1571 TRIE_LIST_NEW( state );
1573 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1574 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1575 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1580 newstate = next_alloc++;
1581 TRIE_LIST_PUSH( state, charid, newstate );
1586 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1590 TRIE_HANDLE_WORD(state);
1592 } /* end second pass */
1594 /* next alloc is the NEXT state to be allocated */
1595 trie->statecount = next_alloc;
1596 trie->states = (reg_trie_state *)
1597 PerlMemShared_realloc( trie->states,
1599 * sizeof(reg_trie_state) );
1601 /* and now dump it out before we compress it */
1602 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1603 revcharmap, next_alloc,
1607 trie->trans = (reg_trie_trans *)
1608 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
1615 for( state=1 ; state < next_alloc ; state ++ ) {
1619 DEBUG_TRIE_COMPILE_MORE_r(
1620 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1624 if (trie->states[state].trans.list) {
1625 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1629 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1630 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1631 if ( forid < minid ) {
1633 } else if ( forid > maxid ) {
1637 if ( transcount < tp + maxid - minid + 1) {
1639 trie->trans = (reg_trie_trans *)
1640 PerlMemShared_realloc( trie->trans,
1642 * sizeof(reg_trie_trans) );
1643 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1645 base = trie->uniquecharcount + tp - minid;
1646 if ( maxid == minid ) {
1648 for ( ; zp < tp ; zp++ ) {
1649 if ( ! trie->trans[ zp ].next ) {
1650 base = trie->uniquecharcount + zp - minid;
1651 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1652 trie->trans[ zp ].check = state;
1658 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1659 trie->trans[ tp ].check = state;
1664 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1665 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
1666 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1667 trie->trans[ tid ].check = state;
1669 tp += ( maxid - minid + 1 );
1671 Safefree(trie->states[ state ].trans.list);
1674 DEBUG_TRIE_COMPILE_MORE_r(
1675 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1678 trie->states[ state ].trans.base=base;
1680 trie->lasttrans = tp + 1;
1684 Second Pass -- Flat Table Representation.
1686 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1687 We know that we will need Charcount+1 trans at most to store the data
1688 (one row per char at worst case) So we preallocate both structures
1689 assuming worst case.
1691 We then construct the trie using only the .next slots of the entry
1694 We use the .check field of the first entry of the node temporarily to
1695 make compression both faster and easier by keeping track of how many non
1696 zero fields are in the node.
1698 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1701 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1702 number representing the first entry of the node, and state as a
1703 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1704 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1705 are 2 entrys per node. eg:
1713 The table is internally in the right hand, idx form. However as we also
1714 have to deal with the states array which is indexed by nodenum we have to
1715 use TRIE_NODENUM() to convert.
1718 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1719 "%*sCompiling trie using table compiler\n",
1720 (int)depth * 2 + 2, ""));
1722 trie->trans = (reg_trie_trans *)
1723 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
1724 * trie->uniquecharcount + 1,
1725 sizeof(reg_trie_trans) );
1726 trie->states = (reg_trie_state *)
1727 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1728 sizeof(reg_trie_state) );
1729 next_alloc = trie->uniquecharcount + 1;
1732 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1734 regnode * const noper = NEXTOPER( cur );
1735 const U8 *uc = (U8*)STRING( noper );
1736 const U8 * const e = uc + STR_LEN( noper );
1738 U32 state = 1; /* required init */
1740 U16 charid = 0; /* sanity init */
1741 U32 accept_state = 0; /* sanity init */
1742 U8 *scan = (U8*)NULL; /* sanity init */
1744 STRLEN foldlen = 0; /* required init */
1745 U32 wordlen = 0; /* required init */
1746 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1748 if ( OP(noper) != NOTHING ) {
1749 for ( ; uc < e ; uc += len ) {
1754 charid = trie->charmap[ uvc ];
1756 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1757 charid = svpp ? (U16)SvIV(*svpp) : 0;
1761 if ( !trie->trans[ state + charid ].next ) {
1762 trie->trans[ state + charid ].next = next_alloc;
1763 trie->trans[ state ].check++;
1764 next_alloc += trie->uniquecharcount;
1766 state = trie->trans[ state + charid ].next;
1768 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1770 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1773 accept_state = TRIE_NODENUM( state );
1774 TRIE_HANDLE_WORD(accept_state);
1776 } /* end second pass */
1778 /* and now dump it out before we compress it */
1779 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
1781 next_alloc, depth+1));
1785 * Inplace compress the table.*
1787 For sparse data sets the table constructed by the trie algorithm will
1788 be mostly 0/FAIL transitions or to put it another way mostly empty.
1789 (Note that leaf nodes will not contain any transitions.)
1791 This algorithm compresses the tables by eliminating most such
1792 transitions, at the cost of a modest bit of extra work during lookup:
1794 - Each states[] entry contains a .base field which indicates the
1795 index in the state[] array wheres its transition data is stored.
1797 - If .base is 0 there are no valid transitions from that node.
1799 - If .base is nonzero then charid is added to it to find an entry in
1802 -If trans[states[state].base+charid].check!=state then the
1803 transition is taken to be a 0/Fail transition. Thus if there are fail
1804 transitions at the front of the node then the .base offset will point
1805 somewhere inside the previous nodes data (or maybe even into a node
1806 even earlier), but the .check field determines if the transition is
1810 The following process inplace converts the table to the compressed
1811 table: We first do not compress the root node 1,and mark its all its
1812 .check pointers as 1 and set its .base pointer as 1 as well. This
1813 allows to do a DFA construction from the compressed table later, and
1814 ensures that any .base pointers we calculate later are greater than
1817 - We set 'pos' to indicate the first entry of the second node.
1819 - We then iterate over the columns of the node, finding the first and
1820 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
1821 and set the .check pointers accordingly, and advance pos
1822 appropriately and repreat for the next node. Note that when we copy
1823 the next pointers we have to convert them from the original
1824 NODEIDX form to NODENUM form as the former is not valid post
1827 - If a node has no transitions used we mark its base as 0 and do not
1828 advance the pos pointer.
1830 - If a node only has one transition we use a second pointer into the
1831 structure to fill in allocated fail transitions from other states.
1832 This pointer is independent of the main pointer and scans forward
1833 looking for null transitions that are allocated to a state. When it
1834 finds one it writes the single transition into the "hole". If the
1835 pointer doesnt find one the single transition is appended as normal.
1837 - Once compressed we can Renew/realloc the structures to release the
1840 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
1841 specifically Fig 3.47 and the associated pseudocode.
1845 const U32 laststate = TRIE_NODENUM( next_alloc );
1848 trie->statecount = laststate;
1850 for ( state = 1 ; state < laststate ; state++ ) {
1852 const U32 stateidx = TRIE_NODEIDX( state );
1853 const U32 o_used = trie->trans[ stateidx ].check;
1854 U32 used = trie->trans[ stateidx ].check;
1855 trie->trans[ stateidx ].check = 0;
1857 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
1858 if ( flag || trie->trans[ stateidx + charid ].next ) {
1859 if ( trie->trans[ stateidx + charid ].next ) {
1861 for ( ; zp < pos ; zp++ ) {
1862 if ( ! trie->trans[ zp ].next ) {
1866 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
1867 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
1868 trie->trans[ zp ].check = state;
1869 if ( ++zp > pos ) pos = zp;
1876 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
1878 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
1879 trie->trans[ pos ].check = state;
1884 trie->lasttrans = pos + 1;
1885 trie->states = (reg_trie_state *)
1886 PerlMemShared_realloc( trie->states, laststate
1887 * sizeof(reg_trie_state) );
1888 DEBUG_TRIE_COMPILE_MORE_r(
1889 PerlIO_printf( Perl_debug_log,
1890 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
1891 (int)depth * 2 + 2,"",
1892 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
1895 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
1898 } /* end table compress */
1900 DEBUG_TRIE_COMPILE_MORE_r(
1901 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
1902 (int)depth * 2 + 2, "",
1903 (UV)trie->statecount,
1904 (UV)trie->lasttrans)
1906 /* resize the trans array to remove unused space */
1907 trie->trans = (reg_trie_trans *)
1908 PerlMemShared_realloc( trie->trans, trie->lasttrans
1909 * sizeof(reg_trie_trans) );
1911 /* and now dump out the compressed format */
1912 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
1914 { /* Modify the program and insert the new TRIE node*/
1915 U8 nodetype =(U8)(flags & 0xFF);
1919 regnode *optimize = NULL;
1920 #ifdef RE_TRACK_PATTERN_OFFSETS
1923 U32 mjd_nodelen = 0;
1924 #endif /* RE_TRACK_PATTERN_OFFSETS */
1925 #endif /* DEBUGGING */
1927 This means we convert either the first branch or the first Exact,
1928 depending on whether the thing following (in 'last') is a branch
1929 or not and whther first is the startbranch (ie is it a sub part of
1930 the alternation or is it the whole thing.)
1931 Assuming its a sub part we conver the EXACT otherwise we convert
1932 the whole branch sequence, including the first.
1934 /* Find the node we are going to overwrite */
1935 if ( first != startbranch || OP( last ) == BRANCH ) {
1936 /* branch sub-chain */
1937 NEXT_OFF( first ) = (U16)(last - first);
1938 #ifdef RE_TRACK_PATTERN_OFFSETS
1940 mjd_offset= Node_Offset((convert));
1941 mjd_nodelen= Node_Length((convert));
1944 /* whole branch chain */
1946 #ifdef RE_TRACK_PATTERN_OFFSETS
1949 const regnode *nop = NEXTOPER( convert );
1950 mjd_offset= Node_Offset((nop));
1951 mjd_nodelen= Node_Length((nop));
1955 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
1956 (int)depth * 2 + 2, "",
1957 (UV)mjd_offset, (UV)mjd_nodelen)
1960 /* But first we check to see if there is a common prefix we can
1961 split out as an EXACT and put in front of the TRIE node. */
1962 trie->startstate= 1;
1963 if ( trie->bitmap && !widecharmap && !trie->jump ) {
1965 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
1969 const U32 base = trie->states[ state ].trans.base;
1971 if ( trie->states[state].wordnum )
1974 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1975 if ( ( base + ofs >= trie->uniquecharcount ) &&
1976 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1977 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1979 if ( ++count > 1 ) {
1980 SV **tmp = av_fetch( revcharmap, ofs, 0);
1981 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
1982 if ( state == 1 ) break;
1984 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
1986 PerlIO_printf(Perl_debug_log,
1987 "%*sNew Start State=%"UVuf" Class: [",
1988 (int)depth * 2 + 2, "",
1991 SV ** const tmp = av_fetch( revcharmap, idx, 0);
1992 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
1994 TRIE_BITMAP_SET(trie,*ch);
1996 TRIE_BITMAP_SET(trie, folder[ *ch ]);
1998 PerlIO_printf(Perl_debug_log, (char*)ch)
2002 TRIE_BITMAP_SET(trie,*ch);
2004 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2005 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2011 SV **tmp = av_fetch( revcharmap, idx, 0);
2013 char *ch = SvPV( *tmp, len );
2015 SV *sv=sv_newmortal();
2016 PerlIO_printf( Perl_debug_log,
2017 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2018 (int)depth * 2 + 2, "",
2020 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2021 PL_colors[0], PL_colors[1],
2022 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2023 PERL_PV_ESCAPE_FIRSTCHAR
2028 OP( convert ) = nodetype;
2029 str=STRING(convert);
2032 STR_LEN(convert) += len;
2038 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2044 regnode *n = convert+NODE_SZ_STR(convert);
2045 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2046 trie->startstate = state;
2047 trie->minlen -= (state - 1);
2048 trie->maxlen -= (state - 1);
2050 /* At least the UNICOS C compiler choked on this
2051 * being argument to DEBUG_r(), so let's just have
2054 #ifdef PERL_EXT_RE_BUILD
2060 regnode *fix = convert;
2061 U32 word = trie->wordcount;
2063 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2064 while( ++fix < n ) {
2065 Set_Node_Offset_Length(fix, 0, 0);
2068 SV ** const tmp = av_fetch( trie_words, word, 0 );
2070 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2071 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2073 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2081 NEXT_OFF(convert) = (U16)(tail - convert);
2082 DEBUG_r(optimize= n);
2088 if ( trie->maxlen ) {
2089 NEXT_OFF( convert ) = (U16)(tail - convert);
2090 ARG_SET( convert, data_slot );
2091 /* Store the offset to the first unabsorbed branch in
2092 jump[0], which is otherwise unused by the jump logic.
2093 We use this when dumping a trie and during optimisation. */
2095 trie->jump[0] = (U16)(nextbranch - convert);
2098 if ( !trie->states[trie->startstate].wordnum && trie->bitmap &&
2099 ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2101 OP( convert ) = TRIEC;
2102 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2103 PerlMemShared_free(trie->bitmap);
2106 OP( convert ) = TRIE;
2108 /* store the type in the flags */
2109 convert->flags = nodetype;
2113 + regarglen[ OP( convert ) ];
2115 /* XXX We really should free up the resource in trie now,
2116 as we won't use them - (which resources?) dmq */
2118 /* needed for dumping*/
2119 DEBUG_r(if (optimize) {
2120 regnode *opt = convert;
2122 while ( ++opt < optimize) {
2123 Set_Node_Offset_Length(opt,0,0);
2126 Try to clean up some of the debris left after the
2129 while( optimize < jumper ) {
2130 mjd_nodelen += Node_Length((optimize));
2131 OP( optimize ) = OPTIMIZED;
2132 Set_Node_Offset_Length(optimize,0,0);
2135 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2137 } /* end node insert */
2138 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2140 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2141 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2143 SvREFCNT_dec(revcharmap);
2147 : trie->startstate>1
2153 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2155 /* The Trie is constructed and compressed now so we can build a fail array now if its needed
2157 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2158 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2161 We find the fail state for each state in the trie, this state is the longest proper
2162 suffix of the current states 'word' that is also a proper prefix of another word in our
2163 trie. State 1 represents the word '' and is the thus the default fail state. This allows
2164 the DFA not to have to restart after its tried and failed a word at a given point, it
2165 simply continues as though it had been matching the other word in the first place.
2167 'abcdgu'=~/abcdefg|cdgu/
2168 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
2169 fail, which would bring use to the state representing 'd' in the second word where we would
2170 try 'g' and succeed, prodceding to match 'cdgu'.
2172 /* add a fail transition */
2173 const U32 trie_offset = ARG(source);
2174 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
2176 const U32 ucharcount = trie->uniquecharcount;
2177 const U32 numstates = trie->statecount;
2178 const U32 ubound = trie->lasttrans + ucharcount;
2182 U32 base = trie->states[ 1 ].trans.base;
2185 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2186 GET_RE_DEBUG_FLAGS_DECL;
2188 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
2190 PERL_UNUSED_ARG(depth);
2194 ARG_SET( stclass, data_slot );
2195 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
2196 RExC_rxi->data->data[ data_slot ] = (void*)aho;
2197 aho->trie=trie_offset;
2198 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2199 Copy( trie->states, aho->states, numstates, reg_trie_state );
2200 Newxz( q, numstates, U32);
2201 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
2204 /* initialize fail[0..1] to be 1 so that we always have
2205 a valid final fail state */
2206 fail[ 0 ] = fail[ 1 ] = 1;
2208 for ( charid = 0; charid < ucharcount ; charid++ ) {
2209 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2211 q[ q_write ] = newstate;
2212 /* set to point at the root */
2213 fail[ q[ q_write++ ] ]=1;
2216 while ( q_read < q_write) {
2217 const U32 cur = q[ q_read++ % numstates ];
2218 base = trie->states[ cur ].trans.base;
2220 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2221 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2223 U32 fail_state = cur;
2226 fail_state = fail[ fail_state ];
2227 fail_base = aho->states[ fail_state ].trans.base;
2228 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2230 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2231 fail[ ch_state ] = fail_state;
2232 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2234 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2236 q[ q_write++ % numstates] = ch_state;
2240 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2241 when we fail in state 1, this allows us to use the
2242 charclass scan to find a valid start char. This is based on the principle
2243 that theres a good chance the string being searched contains lots of stuff
2244 that cant be a start char.
2246 fail[ 0 ] = fail[ 1 ] = 0;
2247 DEBUG_TRIE_COMPILE_r({
2248 PerlIO_printf(Perl_debug_log,
2249 "%*sStclass Failtable (%"UVuf" states): 0",
2250 (int)(depth * 2), "", (UV)numstates
2252 for( q_read=1; q_read<numstates; q_read++ ) {
2253 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2255 PerlIO_printf(Perl_debug_log, "\n");
2258 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
2263 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2264 * These need to be revisited when a newer toolchain becomes available.
2266 #if defined(__sparc64__) && defined(__GNUC__)
2267 # if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2268 # undef SPARC64_GCC_WORKAROUND
2269 # define SPARC64_GCC_WORKAROUND 1
2273 #define DEBUG_PEEP(str,scan,depth) \
2274 DEBUG_OPTIMISE_r({if (scan){ \
2275 SV * const mysv=sv_newmortal(); \
2276 regnode *Next = regnext(scan); \
2277 regprop(RExC_rx, mysv, scan); \
2278 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
2279 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2280 Next ? (REG_NODE_NUM(Next)) : 0 ); \
2287 #define JOIN_EXACT(scan,min,flags) \
2288 if (PL_regkind[OP(scan)] == EXACT) \
2289 join_exact(pRExC_state,(scan),(min),(flags),NULL,depth+1)
2292 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan, I32 *min, U32 flags,regnode *val, U32 depth) {
2293 /* Merge several consecutive EXACTish nodes into one. */
2294 regnode *n = regnext(scan);
2296 regnode *next = scan + NODE_SZ_STR(scan);
2300 regnode *stop = scan;
2301 GET_RE_DEBUG_FLAGS_DECL;
2303 PERL_UNUSED_ARG(depth);
2306 PERL_ARGS_ASSERT_JOIN_EXACT;
2307 #ifndef EXPERIMENTAL_INPLACESCAN
2308 PERL_UNUSED_ARG(flags);
2309 PERL_UNUSED_ARG(val);
2311 DEBUG_PEEP("join",scan,depth);
2313 /* Skip NOTHING, merge EXACT*. */
2315 ( PL_regkind[OP(n)] == NOTHING ||
2316 (stringok && (OP(n) == OP(scan))))
2318 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX) {
2320 if (OP(n) == TAIL || n > next)
2322 if (PL_regkind[OP(n)] == NOTHING) {
2323 DEBUG_PEEP("skip:",n,depth);
2324 NEXT_OFF(scan) += NEXT_OFF(n);
2325 next = n + NODE_STEP_REGNODE;
2332 else if (stringok) {
2333 const unsigned int oldl = STR_LEN(scan);
2334 regnode * const nnext = regnext(n);
2336 DEBUG_PEEP("merg",n,depth);
2339 if (oldl + STR_LEN(n) > U8_MAX)
2341 NEXT_OFF(scan) += NEXT_OFF(n);
2342 STR_LEN(scan) += STR_LEN(n);
2343 next = n + NODE_SZ_STR(n);
2344 /* Now we can overwrite *n : */
2345 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2353 #ifdef EXPERIMENTAL_INPLACESCAN
2354 if (flags && !NEXT_OFF(n)) {
2355 DEBUG_PEEP("atch", val, depth);
2356 if (reg_off_by_arg[OP(n)]) {
2357 ARG_SET(n, val - n);
2360 NEXT_OFF(n) = val - n;
2367 if (UTF && ( OP(scan) == EXACTF ) && ( STR_LEN(scan) >= 6 ) ) {
2369 Two problematic code points in Unicode casefolding of EXACT nodes:
2371 U+0390 - GREEK SMALL LETTER IOTA WITH DIALYTIKA AND TONOS
2372 U+03B0 - GREEK SMALL LETTER UPSILON WITH DIALYTIKA AND TONOS
2378 U+03B9 U+0308 U+0301 0xCE 0xB9 0xCC 0x88 0xCC 0x81
2379 U+03C5 U+0308 U+0301 0xCF 0x85 0xCC 0x88 0xCC 0x81
2381 This means that in case-insensitive matching (or "loose matching",
2382 as Unicode calls it), an EXACTF of length six (the UTF-8 encoded byte
2383 length of the above casefolded versions) can match a target string
2384 of length two (the byte length of UTF-8 encoded U+0390 or U+03B0).
2385 This would rather mess up the minimum length computation.
2387 What we'll do is to look for the tail four bytes, and then peek
2388 at the preceding two bytes to see whether we need to decrease
2389 the minimum length by four (six minus two).
2391 Thanks to the design of UTF-8, there cannot be false matches:
2392 A sequence of valid UTF-8 bytes cannot be a subsequence of
2393 another valid sequence of UTF-8 bytes.
2396 char * const s0 = STRING(scan), *s, *t;
2397 char * const s1 = s0 + STR_LEN(scan) - 1;
2398 char * const s2 = s1 - 4;
2399 #ifdef EBCDIC /* RD tunifold greek 0390 and 03B0 */
2400 const char t0[] = "\xaf\x49\xaf\x42";
2402 const char t0[] = "\xcc\x88\xcc\x81";
2404 const char * const t1 = t0 + 3;
2407 s < s2 && (t = ninstr(s, s1, t0, t1));
2410 if (((U8)t[-1] == 0x68 && (U8)t[-2] == 0xB4) ||
2411 ((U8)t[-1] == 0x46 && (U8)t[-2] == 0xB5))
2413 if (((U8)t[-1] == 0xB9 && (U8)t[-2] == 0xCE) ||
2414 ((U8)t[-1] == 0x85 && (U8)t[-2] == 0xCF))
2422 n = scan + NODE_SZ_STR(scan);
2424 if (PL_regkind[OP(n)] != NOTHING || OP(n) == NOTHING) {
2431 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2435 /* REx optimizer. Converts nodes into quickier variants "in place".
2436 Finds fixed substrings. */
2438 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
2439 to the position after last scanned or to NULL. */
2441 #define INIT_AND_WITHP \
2442 assert(!and_withp); \
2443 Newx(and_withp,1,struct regnode_charclass_class); \
2444 SAVEFREEPV(and_withp)
2446 /* this is a chain of data about sub patterns we are processing that
2447 need to be handled seperately/specially in study_chunk. Its so
2448 we can simulate recursion without losing state. */
2450 typedef struct scan_frame {
2451 regnode *last; /* last node to process in this frame */
2452 regnode *next; /* next node to process when last is reached */
2453 struct scan_frame *prev; /*previous frame*/
2454 I32 stop; /* what stopparen do we use */
2458 #define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2460 #define CASE_SYNST_FNC(nAmE) \
2462 if (flags & SCF_DO_STCLASS_AND) { \
2463 for (value = 0; value < 256; value++) \
2464 if (!is_ ## nAmE ## _cp(value)) \
2465 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2468 for (value = 0; value < 256; value++) \
2469 if (is_ ## nAmE ## _cp(value)) \
2470 ANYOF_BITMAP_SET(data->start_class, value); \
2474 if (flags & SCF_DO_STCLASS_AND) { \
2475 for (value = 0; value < 256; value++) \
2476 if (is_ ## nAmE ## _cp(value)) \
2477 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2480 for (value = 0; value < 256; value++) \
2481 if (!is_ ## nAmE ## _cp(value)) \
2482 ANYOF_BITMAP_SET(data->start_class, value); \
2489 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
2490 I32 *minlenp, I32 *deltap,
2495 struct regnode_charclass_class *and_withp,
2496 U32 flags, U32 depth)
2497 /* scanp: Start here (read-write). */
2498 /* deltap: Write maxlen-minlen here. */
2499 /* last: Stop before this one. */
2500 /* data: string data about the pattern */
2501 /* stopparen: treat close N as END */
2502 /* recursed: which subroutines have we recursed into */
2503 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
2506 I32 min = 0, pars = 0, code;
2507 regnode *scan = *scanp, *next;
2509 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
2510 int is_inf_internal = 0; /* The studied chunk is infinite */
2511 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
2512 scan_data_t data_fake;
2513 SV *re_trie_maxbuff = NULL;
2514 regnode *first_non_open = scan;
2515 I32 stopmin = I32_MAX;
2516 scan_frame *frame = NULL;
2517 GET_RE_DEBUG_FLAGS_DECL;
2519 PERL_ARGS_ASSERT_STUDY_CHUNK;
2522 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
2526 while (first_non_open && OP(first_non_open) == OPEN)
2527 first_non_open=regnext(first_non_open);
2532 while ( scan && OP(scan) != END && scan < last ){
2533 /* Peephole optimizer: */
2534 DEBUG_STUDYDATA("Peep:", data,depth);
2535 DEBUG_PEEP("Peep",scan,depth);
2536 JOIN_EXACT(scan,&min,0);
2538 /* Follow the next-chain of the current node and optimize
2539 away all the NOTHINGs from it. */
2540 if (OP(scan) != CURLYX) {
2541 const int max = (reg_off_by_arg[OP(scan)]
2543 /* I32 may be smaller than U16 on CRAYs! */
2544 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
2545 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
2549 /* Skip NOTHING and LONGJMP. */
2550 while ((n = regnext(n))
2551 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
2552 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
2553 && off + noff < max)
2555 if (reg_off_by_arg[OP(scan)])
2558 NEXT_OFF(scan) = off;
2563 /* The principal pseudo-switch. Cannot be a switch, since we
2564 look into several different things. */
2565 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
2566 || OP(scan) == IFTHEN) {
2567 next = regnext(scan);
2569 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
2571 if (OP(next) == code || code == IFTHEN) {
2572 /* NOTE - There is similar code to this block below for handling
2573 TRIE nodes on a re-study. If you change stuff here check there
2575 I32 max1 = 0, min1 = I32_MAX, num = 0;
2576 struct regnode_charclass_class accum;
2577 regnode * const startbranch=scan;
2579 if (flags & SCF_DO_SUBSTR)
2580 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
2581 if (flags & SCF_DO_STCLASS)
2582 cl_init_zero(pRExC_state, &accum);
2584 while (OP(scan) == code) {
2585 I32 deltanext, minnext, f = 0, fake;
2586 struct regnode_charclass_class this_class;
2589 data_fake.flags = 0;
2591 data_fake.whilem_c = data->whilem_c;
2592 data_fake.last_closep = data->last_closep;
2595 data_fake.last_closep = &fake;
2597 data_fake.pos_delta = delta;
2598 next = regnext(scan);
2599 scan = NEXTOPER(scan);
2601 scan = NEXTOPER(scan);
2602 if (flags & SCF_DO_STCLASS) {
2603 cl_init(pRExC_state, &this_class);
2604 data_fake.start_class = &this_class;
2605 f = SCF_DO_STCLASS_AND;
2607 if (flags & SCF_WHILEM_VISITED_POS)
2608 f |= SCF_WHILEM_VISITED_POS;
2610 /* we suppose the run is continuous, last=next...*/
2611 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
2613 stopparen, recursed, NULL, f,depth+1);
2616 if (max1 < minnext + deltanext)
2617 max1 = minnext + deltanext;
2618 if (deltanext == I32_MAX)
2619 is_inf = is_inf_internal = 1;
2621 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
2623 if (data_fake.flags & SCF_SEEN_ACCEPT) {
2624 if ( stopmin > minnext)
2625 stopmin = min + min1;
2626 flags &= ~SCF_DO_SUBSTR;
2628 data->flags |= SCF_SEEN_ACCEPT;
2631 if (data_fake.flags & SF_HAS_EVAL)
2632 data->flags |= SF_HAS_EVAL;
2633 data->whilem_c = data_fake.whilem_c;
2635 if (flags & SCF_DO_STCLASS)
2636 cl_or(pRExC_state, &accum, &this_class);
2638 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
2640 if (flags & SCF_DO_SUBSTR) {
2641 data->pos_min += min1;
2642 data->pos_delta += max1 - min1;
2643 if (max1 != min1 || is_inf)
2644 data->longest = &(data->longest_float);
2647 delta += max1 - min1;
2648 if (flags & SCF_DO_STCLASS_OR) {
2649 cl_or(pRExC_state, data->start_class, &accum);
2651 cl_and(data->start_class, and_withp);
2652 flags &= ~SCF_DO_STCLASS;
2655 else if (flags & SCF_DO_STCLASS_AND) {
2657 cl_and(data->start_class, &accum);
2658 flags &= ~SCF_DO_STCLASS;
2661 /* Switch to OR mode: cache the old value of
2662 * data->start_class */
2664 StructCopy(data->start_class, and_withp,
2665 struct regnode_charclass_class);
2666 flags &= ~SCF_DO_STCLASS_AND;
2667 StructCopy(&accum, data->start_class,
2668 struct regnode_charclass_class);
2669 flags |= SCF_DO_STCLASS_OR;
2670 data->start_class->flags |= ANYOF_EOS;
2674 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
2677 Assuming this was/is a branch we are dealing with: 'scan' now
2678 points at the item that follows the branch sequence, whatever
2679 it is. We now start at the beginning of the sequence and look
2686 which would be constructed from a pattern like /A|LIST|OF|WORDS/
2688 If we can find such a subseqence we need to turn the first
2689 element into a trie and then add the subsequent branch exact
2690 strings to the trie.
2694 1. patterns where the whole set of branch can be converted.
2696 2. patterns where only a subset can be converted.
2698 In case 1 we can replace the whole set with a single regop
2699 for the trie. In case 2 we need to keep the start and end
2702 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
2703 becomes BRANCH TRIE; BRANCH X;
2705 There is an additional case, that being where there is a
2706 common prefix, which gets split out into an EXACT like node
2707 preceding the TRIE node.
2709 If x(1..n)==tail then we can do a simple trie, if not we make
2710 a "jump" trie, such that when we match the appropriate word
2711 we "jump" to the appopriate tail node. Essentailly we turn
2712 a nested if into a case structure of sorts.
2717 if (!re_trie_maxbuff) {
2718 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2719 if (!SvIOK(re_trie_maxbuff))
2720 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2722 if ( SvIV(re_trie_maxbuff)>=0 ) {
2724 regnode *first = (regnode *)NULL;
2725 regnode *last = (regnode *)NULL;
2726 regnode *tail = scan;
2731 SV * const mysv = sv_newmortal(); /* for dumping */
2733 /* var tail is used because there may be a TAIL
2734 regop in the way. Ie, the exacts will point to the
2735 thing following the TAIL, but the last branch will
2736 point at the TAIL. So we advance tail. If we
2737 have nested (?:) we may have to move through several
2741 while ( OP( tail ) == TAIL ) {
2742 /* this is the TAIL generated by (?:) */
2743 tail = regnext( tail );
2748 regprop(RExC_rx, mysv, tail );
2749 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
2750 (int)depth * 2 + 2, "",
2751 "Looking for TRIE'able sequences. Tail node is: ",
2752 SvPV_nolen_const( mysv )
2758 step through the branches, cur represents each
2759 branch, noper is the first thing to be matched
2760 as part of that branch and noper_next is the
2761 regnext() of that node. if noper is an EXACT
2762 and noper_next is the same as scan (our current
2763 position in the regex) then the EXACT branch is
2764 a possible optimization target. Once we have
2765 two or more consequetive such branches we can
2766 create a trie of the EXACT's contents and stich
2767 it in place. If the sequence represents all of
2768 the branches we eliminate the whole thing and
2769 replace it with a single TRIE. If it is a
2770 subsequence then we need to stitch it in. This
2771 means the first branch has to remain, and needs
2772 to be repointed at the item on the branch chain
2773 following the last branch optimized. This could
2774 be either a BRANCH, in which case the
2775 subsequence is internal, or it could be the
2776 item following the branch sequence in which
2777 case the subsequence is at the end.
2781 /* dont use tail as the end marker for this traverse */
2782 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
2783 regnode * const noper = NEXTOPER( cur );
2784 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
2785 regnode * const noper_next = regnext( noper );
2789 regprop(RExC_rx, mysv, cur);
2790 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
2791 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
2793 regprop(RExC_rx, mysv, noper);
2794 PerlIO_printf( Perl_debug_log, " -> %s",
2795 SvPV_nolen_const(mysv));
2798 regprop(RExC_rx, mysv, noper_next );
2799 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
2800 SvPV_nolen_const(mysv));
2802 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d)\n",
2803 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur) );
2805 if ( (((first && optype!=NOTHING) ? OP( noper ) == optype
2806 : PL_regkind[ OP( noper ) ] == EXACT )
2807 || OP(noper) == NOTHING )
2809 && noper_next == tail
2814 if ( !first || optype == NOTHING ) {
2815 if (!first) first = cur;
2816 optype = OP( noper );
2822 Currently we assume that the trie can handle unicode and ascii
2823 matches fold cased matches. If this proves true then the following
2824 define will prevent tries in this situation.
2826 #define TRIE_TYPE_IS_SAFE (UTF || optype==EXACT)
2828 #define TRIE_TYPE_IS_SAFE 1
2829 if ( last && TRIE_TYPE_IS_SAFE ) {
2830 make_trie( pRExC_state,
2831 startbranch, first, cur, tail, count,
2834 if ( PL_regkind[ OP( noper ) ] == EXACT
2836 && noper_next == tail
2841 optype = OP( noper );
2851 regprop(RExC_rx, mysv, cur);
2852 PerlIO_printf( Perl_debug_log,
2853 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
2854 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
2858 if ( last && TRIE_TYPE_IS_SAFE ) {
2859 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, optype, depth+1 );
2860 #ifdef TRIE_STUDY_OPT
2861 if ( ((made == MADE_EXACT_TRIE &&
2862 startbranch == first)
2863 || ( first_non_open == first )) &&
2865 flags |= SCF_TRIE_RESTUDY;
2866 if ( startbranch == first
2869 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
2879 else if ( code == BRANCHJ ) { /* single branch is optimized. */
2880 scan = NEXTOPER(NEXTOPER(scan));
2881 } else /* single branch is optimized. */
2882 scan = NEXTOPER(scan);
2884 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
2885 scan_frame *newframe = NULL;
2890 if (OP(scan) != SUSPEND) {
2891 /* set the pointer */
2892 if (OP(scan) == GOSUB) {
2894 RExC_recurse[ARG2L(scan)] = scan;
2895 start = RExC_open_parens[paren-1];
2896 end = RExC_close_parens[paren-1];
2899 start = RExC_rxi->program + 1;
2903 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
2904 SAVEFREEPV(recursed);
2906 if (!PAREN_TEST(recursed,paren+1)) {
2907 PAREN_SET(recursed,paren+1);
2908 Newx(newframe,1,scan_frame);
2910 if (flags & SCF_DO_SUBSTR) {
2911 SCAN_COMMIT(pRExC_state,data,minlenp);
2912 data->longest = &(data->longest_float);
2914 is_inf = is_inf_internal = 1;
2915 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
2916 cl_anything(pRExC_state, data->start_class);
2917 flags &= ~SCF_DO_STCLASS;
2920 Newx(newframe,1,scan_frame);
2923 end = regnext(scan);
2928 SAVEFREEPV(newframe);
2929 newframe->next = regnext(scan);
2930 newframe->last = last;
2931 newframe->stop = stopparen;
2932 newframe->prev = frame;
2942 else if (OP(scan) == EXACT) {
2943 I32 l = STR_LEN(scan);
2946 const U8 * const s = (U8*)STRING(scan);
2947 l = utf8_length(s, s + l);
2948 uc = utf8_to_uvchr(s, NULL);
2950 uc = *((U8*)STRING(scan));
2953 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
2954 /* The code below prefers earlier match for fixed
2955 offset, later match for variable offset. */
2956 if (data->last_end == -1) { /* Update the start info. */
2957 data->last_start_min = data->pos_min;
2958 data->last_start_max = is_inf
2959 ? I32_MAX : data->pos_min + data->pos_delta;
2961 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
2963 SvUTF8_on(data->last_found);
2965 SV * const sv = data->last_found;
2966 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
2967 mg_find(sv, PERL_MAGIC_utf8) : NULL;
2968 if (mg && mg->mg_len >= 0)
2969 mg->mg_len += utf8_length((U8*)STRING(scan),
2970 (U8*)STRING(scan)+STR_LEN(scan));
2972 data->last_end = data->pos_min + l;
2973 data->pos_min += l; /* As in the first entry. */
2974 data->flags &= ~SF_BEFORE_EOL;
2976 if (flags & SCF_DO_STCLASS_AND) {
2977 /* Check whether it is compatible with what we know already! */
2981 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
2982 && !ANYOF_BITMAP_TEST(data->start_class, uc)
2983 && (!(data->start_class->flags & ANYOF_FOLD)
2984 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold[uc])))
2987 ANYOF_CLASS_ZERO(data->start_class);
2988 ANYOF_BITMAP_ZERO(data->start_class);
2990 ANYOF_BITMAP_SET(data->start_class, uc);
2991 data->start_class->flags &= ~ANYOF_EOS;
2993 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
2995 else if (flags & SCF_DO_STCLASS_OR) {
2996 /* false positive possible if the class is case-folded */
2998 ANYOF_BITMAP_SET(data->start_class, uc);
3000 data->start_class->flags |= ANYOF_UNICODE_ALL;
3001 data->start_class->flags &= ~ANYOF_EOS;
3002 cl_and(data->start_class, and_withp);
3004 flags &= ~SCF_DO_STCLASS;
3006 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
3007 I32 l = STR_LEN(scan);
3008 UV uc = *((U8*)STRING(scan));
3010 /* Search for fixed substrings supports EXACT only. */
3011 if (flags & SCF_DO_SUBSTR) {
3013 SCAN_COMMIT(pRExC_state, data, minlenp);
3016 const U8 * const s = (U8 *)STRING(scan);
3017 l = utf8_length(s, s + l);
3018 uc = utf8_to_uvchr(s, NULL);
3021 if (flags & SCF_DO_SUBSTR)
3023 if (flags & SCF_DO_STCLASS_AND) {
3024 /* Check whether it is compatible with what we know already! */
3028 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3029 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3030 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold[uc])))
3032 ANYOF_CLASS_ZERO(data->start_class);
3033 ANYOF_BITMAP_ZERO(data->start_class);
3035 ANYOF_BITMAP_SET(data->start_class, uc);
3036 data->start_class->flags &= ~ANYOF_EOS;
3037 data->start_class->flags |= ANYOF_FOLD;
3038 if (OP(scan) == EXACTFL)
3039 data->start_class->flags |= ANYOF_LOCALE;
3042 else if (flags & SCF_DO_STCLASS_OR) {
3043 if (data->start_class->flags & ANYOF_FOLD) {
3044 /* false positive possible if the class is case-folded.
3045 Assume that the locale settings are the same... */
3047 ANYOF_BITMAP_SET(data->start_class, uc);
3048 data->start_class->flags &= ~ANYOF_EOS;
3050 cl_and(data->start_class, and_withp);
3052 flags &= ~SCF_DO_STCLASS;
3054 else if (strchr((const char*)PL_varies,OP(scan))) {
3055 I32 mincount, maxcount, minnext, deltanext, fl = 0;
3056 I32 f = flags, pos_before = 0;
3057 regnode * const oscan = scan;
3058 struct regnode_charclass_class this_class;
3059 struct regnode_charclass_class *oclass = NULL;
3060 I32 next_is_eval = 0;
3062 switch (PL_regkind[OP(scan)]) {
3063 case WHILEM: /* End of (?:...)* . */
3064 scan = NEXTOPER(scan);
3067 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
3068 next = NEXTOPER(scan);
3069 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
3071 maxcount = REG_INFTY;
3072 next = regnext(scan);
3073 scan = NEXTOPER(scan);
3077 if (flags & SCF_DO_SUBSTR)
3082 if (flags & SCF_DO_STCLASS) {
3084 maxcount = REG_INFTY;
3085 next = regnext(scan);
3086 scan = NEXTOPER(scan);
3089 is_inf = is_inf_internal = 1;
3090 scan = regnext(scan);
3091 if (flags & SCF_DO_SUBSTR) {
3092 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot extend fixed substrings */
3093 data->longest = &(data->longest_float);
3095 goto optimize_curly_tail;
3097 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
3098 && (scan->flags == stopparen))
3103 mincount = ARG1(scan);
3104 maxcount = ARG2(scan);
3106 next = regnext(scan);
3107 if (OP(scan) == CURLYX) {
3108 I32 lp = (data ? *(data->last_closep) : 0);
3109 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
3111 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
3112 next_is_eval = (OP(scan) == EVAL);
3114 if (flags & SCF_DO_SUBSTR) {
3115 if (mincount == 0) SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot extend fixed substrings */
3116 pos_before = data->pos_min;
3120 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
3122 data->flags |= SF_IS_INF;
3124 if (flags & SCF_DO_STCLASS) {
3125 cl_init(pRExC_state, &this_class);
3126 oclass = data->start_class;
3127 data->start_class = &this_class;
3128 f |= SCF_DO_STCLASS_AND;
3129 f &= ~SCF_DO_STCLASS_OR;
3131 /* These are the cases when once a subexpression
3132 fails at a particular position, it cannot succeed
3133 even after backtracking at the enclosing scope.
3135 XXXX what if minimal match and we are at the
3136 initial run of {n,m}? */
3137 if ((mincount != maxcount - 1) && (maxcount != REG_INFTY))
3138 f &= ~SCF_WHILEM_VISITED_POS;
3140 /* This will finish on WHILEM, setting scan, or on NULL: */
3141 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3142 last, data, stopparen, recursed, NULL,
3144 ? (f & ~SCF_DO_SUBSTR) : f),depth+1);
3146 if (flags & SCF_DO_STCLASS)
3147 data->start_class = oclass;
3148 if (mincount == 0 || minnext == 0) {
3149 if (flags & SCF_DO_STCLASS_OR) {
3150 cl_or(pRExC_state, data->start_class, &this_class);
3152 else if (flags & SCF_DO_STCLASS_AND) {
3153 /* Switch to OR mode: cache the old value of
3154 * data->start_class */
3156 StructCopy(data->start_class, and_withp,
3157 struct regnode_charclass_class);
3158 flags &= ~SCF_DO_STCLASS_AND;
3159 StructCopy(&this_class, data->start_class,
3160 struct regnode_charclass_class);
3161 flags |= SCF_DO_STCLASS_OR;
3162 data->start_class->flags |= ANYOF_EOS;
3164 } else { /* Non-zero len */
3165 if (flags & SCF_DO_STCLASS_OR) {
3166 cl_or(pRExC_state, data->start_class, &this_class);
3167 cl_and(data->start_class, and_withp);
3169 else if (flags & SCF_DO_STCLASS_AND)
3170 cl_and(data->start_class, &this_class);
3171 flags &= ~SCF_DO_STCLASS;
3173 if (!scan) /* It was not CURLYX, but CURLY. */
3175 if ( /* ? quantifier ok, except for (?{ ... }) */
3176 (next_is_eval || !(mincount == 0 && maxcount == 1))
3177 && (minnext == 0) && (deltanext == 0)
3178 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
3179 && maxcount <= REG_INFTY/3 /* Complement check for big count */
3180 && ckWARN(WARN_REGEXP))
3183 "Quantifier unexpected on zero-length expression");
3186 min += minnext * mincount;
3187 is_inf_internal |= ((maxcount == REG_INFTY
3188 && (minnext + deltanext) > 0)
3189 || deltanext == I32_MAX);
3190 is_inf |= is_inf_internal;
3191 delta += (minnext + deltanext) * maxcount - minnext * mincount;
3193 /* Try powerful optimization CURLYX => CURLYN. */
3194 if ( OP(oscan) == CURLYX && data
3195 && data->flags & SF_IN_PAR
3196 && !(data->flags & SF_HAS_EVAL)
3197 && !deltanext && minnext == 1 ) {
3198 /* Try to optimize to CURLYN. */
3199 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
3200 regnode * const nxt1 = nxt;
3207 if (!strchr((const char*)PL_simple,OP(nxt))
3208 && !(PL_regkind[OP(nxt)] == EXACT
3209 && STR_LEN(nxt) == 1))
3215 if (OP(nxt) != CLOSE)
3217 if (RExC_open_parens) {
3218 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
3219 RExC_close_parens[ARG(nxt1)-1]=nxt+2; /*close->while*/
3221 /* Now we know that nxt2 is the only contents: */
3222 oscan->flags = (U8)ARG(nxt);
3224 OP(nxt1) = NOTHING; /* was OPEN. */
3227 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
3228 NEXT_OFF(nxt1+ 1) = 0; /* just for consistancy. */
3229 NEXT_OFF(nxt2) = 0; /* just for consistancy with CURLY. */
3230 OP(nxt) = OPTIMIZED; /* was CLOSE. */
3231 OP(nxt + 1) = OPTIMIZED; /* was count. */
3232 NEXT_OFF(nxt+ 1) = 0; /* just for consistancy. */
3237 /* Try optimization CURLYX => CURLYM. */
3238 if ( OP(oscan) == CURLYX && data
3239 && !(data->flags & SF_HAS_PAR)
3240 && !(data->flags & SF_HAS_EVAL)
3241 && !deltanext /* atom is fixed width */
3242 && minnext != 0 /* CURLYM can't handle zero width */
3244 /* XXXX How to optimize if data == 0? */
3245 /* Optimize to a simpler form. */
3246 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
3250 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
3251 && (OP(nxt2) != WHILEM))
3253 OP(nxt2) = SUCCEED; /* Whas WHILEM */
3254 /* Need to optimize away parenths. */
3255 if (data->flags & SF_IN_PAR) {
3256 /* Set the parenth number. */
3257 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
3259 if (OP(nxt) != CLOSE)
3260 FAIL("Panic opt close");
3261 oscan->flags = (U8)ARG(nxt);
3262 if (RExC_open_parens) {
3263 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
3264 RExC_close_parens[ARG(nxt1)-1]=nxt2+1; /*close->NOTHING*/
3266 OP(nxt1) = OPTIMIZED; /* was OPEN. */
3267 OP(nxt) = OPTIMIZED; /* was CLOSE. */
3270 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
3271 OP(nxt + 1) = OPTIMIZED; /* was count. */
3272 NEXT_OFF(nxt1 + 1) = 0; /* just for consistancy. */
3273 NEXT_OFF(nxt + 1) = 0; /* just for consistancy. */
3276 while ( nxt1 && (OP(nxt1) != WHILEM)) {
3277 regnode *nnxt = regnext(nxt1);
3280 if (reg_off_by_arg[OP(nxt1)])
3281 ARG_SET(nxt1, nxt2 - nxt1);
3282 else if (nxt2 - nxt1 < U16_MAX)
3283 NEXT_OFF(nxt1) = nxt2 - nxt1;
3285 OP(nxt) = NOTHING; /* Cannot beautify */
3290 /* Optimize again: */
3291 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
3292 NULL, stopparen, recursed, NULL, 0,depth+1);
3297 else if ((OP(oscan) == CURLYX)
3298 && (flags & SCF_WHILEM_VISITED_POS)
3299 /* See the comment on a similar expression above.
3300 However, this time it not a subexpression
3301 we care about, but the expression itself. */
3302 && (maxcount == REG_INFTY)
3303 && data && ++data->whilem_c < 16) {
3304 /* This stays as CURLYX, we can put the count/of pair. */
3305 /* Find WHILEM (as in regexec.c) */
3306 regnode *nxt = oscan + NEXT_OFF(oscan);
3308 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
3310 PREVOPER(nxt)->flags = (U8)(data->whilem_c
3311 | (RExC_whilem_seen << 4)); /* On WHILEM */
3313 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
3315 if (flags & SCF_DO_SUBSTR) {
3316 SV *last_str = NULL;
3317 int counted = mincount != 0;
3319 if (data->last_end > 0 && mincount != 0) { /* Ends with a string. */
3320 #if defined(SPARC64_GCC_WORKAROUND)
3323 const char *s = NULL;
3326 if (pos_before >= data->last_start_min)
3329 b = data->last_start_min;
3332 s = SvPV_const(data->last_found, l);
3333 old = b - data->last_start_min;
3336 I32 b = pos_before >= data->last_start_min
3337 ? pos_before : data->last_start_min;
3339 const char * const s = SvPV_const(data->last_found, l);
3340 I32 old = b - data->last_start_min;
3344 old = utf8_hop((U8*)s, old) - (U8*)s;
3347 /* Get the added string: */
3348 last_str = newSVpvn_utf8(s + old, l, UTF);
3349 if (deltanext == 0 && pos_before == b) {
3350 /* What was added is a constant string */
3352 SvGROW(last_str, (mincount * l) + 1);
3353 repeatcpy(SvPVX(last_str) + l,
3354 SvPVX_const(last_str), l, mincount - 1);
3355 SvCUR_set(last_str, SvCUR(last_str) * mincount);
3356 /* Add additional parts. */
3357 SvCUR_set(data->last_found,
3358 SvCUR(data->last_found) - l);
3359 sv_catsv(data->last_found, last_str);
3361 SV * sv = data->last_found;
3363 SvUTF8(sv) && SvMAGICAL(sv) ?
3364 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3365 if (mg && mg->mg_len >= 0)
3366 mg->mg_len += CHR_SVLEN(last_str) - l;
3368 data->last_end += l * (mincount - 1);
3371 /* start offset must point into the last copy */
3372 data->last_start_min += minnext * (mincount - 1);
3373 data->last_start_max += is_inf ? I32_MAX
3374 : (maxcount - 1) * (minnext + data->pos_delta);
3377 /* It is counted once already... */
3378 data->pos_min += minnext * (mincount - counted);
3379 data->pos_delta += - counted * deltanext +
3380 (minnext + deltanext) * maxcount - minnext * mincount;
3381 if (mincount != maxcount) {
3382 /* Cannot extend fixed substrings found inside
3384 SCAN_COMMIT(pRExC_state,data,minlenp);
3385 if (mincount && last_str) {
3386 SV * const sv = data->last_found;
3387 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3388 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3392 sv_setsv(sv, last_str);
3393 data->last_end = data->pos_min;
3394 data->last_start_min =
3395 data->pos_min - CHR_SVLEN(last_str);
3396 data->last_start_max = is_inf
3398 : data->pos_min + data->pos_delta
3399 - CHR_SVLEN(last_str);
3401 data->longest = &(data->longest_float);
3403 SvREFCNT_dec(last_str);
3405 if (data && (fl & SF_HAS_EVAL))
3406 data->flags |= SF_HAS_EVAL;
3407 optimize_curly_tail:
3408 if (OP(oscan) != CURLYX) {
3409 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
3411 NEXT_OFF(oscan) += NEXT_OFF(next);
3414 default: /* REF and CLUMP only? */
3415 if (flags & SCF_DO_SUBSTR) {
3416 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
3417 data->longest = &(data->longest_float);
3419 is_inf = is_inf_internal = 1;
3420 if (flags & SCF_DO_STCLASS_OR)
3421 cl_anything(pRExC_state, data->start_class);
3422 flags &= ~SCF_DO_STCLASS;
3426 else if (OP(scan) == LNBREAK) {
3427 if (flags & SCF_DO_STCLASS) {
3429 data->start_class->flags &= ~ANYOF_EOS; /* No match on empty */
3430 if (flags & SCF_DO_STCLASS_AND) {
3431 for (value = 0; value < 256; value++)
3432 if (!is_VERTWS_cp(value))
3433 ANYOF_BITMAP_CLEAR(data->start_class, value);
3436 for (value = 0; value < 256; value++)
3437 if (is_VERTWS_cp(value))
3438 ANYOF_BITMAP_SET(data->start_class, value);
3440 if (flags & SCF_DO_STCLASS_OR)
3441 cl_and(data->start_class, and_withp);
3442 flags &= ~SCF_DO_STCLASS;
3446 if (flags & SCF_DO_SUBSTR) {
3447 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
3449 data->pos_delta += 1;
3450 data->longest = &(data->longest_float);
3454 else if (OP(scan) == FOLDCHAR) {
3455 int d = ARG(scan)==0xDF ? 1 : 2;
3456 flags &= ~SCF_DO_STCLASS;
3459 if (flags & SCF_DO_SUBSTR) {
3460 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
3462 data->pos_delta += d;
3463 data->longest = &(data->longest_float);
3466 else if (strchr((const char*)PL_simple,OP(scan))) {
3469 if (flags & SCF_DO_SUBSTR) {
3470 SCAN_COMMIT(pRExC_state,data,minlenp);
3474 if (flags & SCF_DO_STCLASS) {
3475 data->start_class->flags &= ~ANYOF_EOS; /* No match on empty */
3477 /* Some of the logic below assumes that switching
3478 locale on will only add false positives. */
3479 switch (PL_regkind[OP(scan)]) {
3483 /* Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d", OP(scan)); */
3484 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3485 cl_anything(pRExC_state, data->start_class);
3488 if (OP(scan) == SANY)
3490 if (flags & SCF_DO_STCLASS_OR) { /* Everything but \n */
3491 value = (ANYOF_BITMAP_TEST(data->start_class,'\n')
3492 || (data->start_class->flags & ANYOF_CLASS));
3493 cl_anything(pRExC_state, data->start_class);
3495 if (flags & SCF_DO_STCLASS_AND || !value)
3496 ANYOF_BITMAP_CLEAR(data->start_class,'\n');
3499 if (flags & SCF_DO_STCLASS_AND)
3500 cl_and(data->start_class,
3501 (struct regnode_charclass_class*)scan);
3503 cl_or(pRExC_state, data->start_class,
3504 (struct regnode_charclass_class*)scan);
3507 if (flags & SCF_DO_STCLASS_AND) {
3508 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3509 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NALNUM);
3510 for (value = 0; value < 256; value++)
3511 if (!isALNUM(value))
3512 ANYOF_BITMAP_CLEAR(data->start_class, value);
3516 if (data->start_class->flags & ANYOF_LOCALE)
3517 ANYOF_CLASS_SET(data->start_class,ANYOF_ALNUM);
3519 for (value = 0; value < 256; value++)
3521 ANYOF_BITMAP_SET(data->start_class, value);
3526 if (flags & SCF_DO_STCLASS_AND) {
3527 if (data->start_class->flags & ANYOF_LOCALE)
3528 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NALNUM);
3531 ANYOF_CLASS_SET(data->start_class,ANYOF_ALNUM);
3532 data->start_class->flags |= ANYOF_LOCALE;
3536 if (flags & SCF_DO_STCLASS_AND) {
3537 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3538 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_ALNUM);
3539 for (value = 0; value < 256; value++)
3541 ANYOF_BITMAP_CLEAR(data->start_class, value);
3545 if (data->start_class->flags & ANYOF_LOCALE)
3546 ANYOF_CLASS_SET(data->start_class,ANYOF_NALNUM);
3548 for (value = 0; value < 256; value++)
3549 if (!isALNUM(value))
3550 ANYOF_BITMAP_SET(data->start_class, value);
3555 if (flags & SCF_DO_STCLASS_AND) {
3556 if (data->start_class->flags & ANYOF_LOCALE)
3557 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_ALNUM);
3560 data->start_class->flags |= ANYOF_LOCALE;
3561 ANYOF_CLASS_SET(data->start_class,ANYOF_NALNUM);
3565 if (flags & SCF_DO_STCLASS_AND) {
3566 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3567 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NSPACE);
3568 for (value = 0; value < 256; value++)
3569 if (!isSPACE(value))
3570 ANYOF_BITMAP_CLEAR(data->start_class, value);
3574 if (data->start_class->flags & ANYOF_LOCALE)
3575 ANYOF_CLASS_SET(data->start_class,ANYOF_SPACE);
3577 for (value = 0; value < 256; value++)
3579 ANYOF_BITMAP_SET(data->start_class, value);
3584 if (flags & SCF_DO_STCLASS_AND) {
3585 if (data->start_class->flags & ANYOF_LOCALE)
3586 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NSPACE);
3589 data->start_class->flags |= ANYOF_LOCALE;
3590 ANYOF_CLASS_SET(data->start_class,ANYOF_SPACE);
3594 if (flags & SCF_DO_STCLASS_AND) {
3595 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3596 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_SPACE);
3597 for (value = 0; value < 256; value++)
3599 ANYOF_BITMAP_CLEAR(data->start_class, value);
3603 if (data->start_class->flags & ANYOF_LOCALE)
3604 ANYOF_CLASS_SET(data->start_class,ANYOF_NSPACE);
3606 for (value = 0; value < 256; value++)
3607 if (!isSPACE(value))
3608 ANYOF_BITMAP_SET(data->start_class, value);
3613 if (flags & SCF_DO_STCLASS_AND) {
3614 if (data->start_class->flags & ANYOF_LOCALE) {
3615 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_SPACE);
3616 for (value = 0; value < 256; value++)
3617 if (!isSPACE(value))
3618 ANYOF_BITMAP_CLEAR(data->start_class, value);
3622 data->start_class->flags |= ANYOF_LOCALE;
3623 ANYOF_CLASS_SET(data->start_class,ANYOF_NSPACE);
3627 if (flags & SCF_DO_STCLASS_AND) {
3628 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NDIGIT);
3629 for (value = 0; value < 256; value++)
3630 if (!isDIGIT(value))
3631 ANYOF_BITMAP_CLEAR(data->start_class, value);
3634 if (data->start_class->flags & ANYOF_LOCALE)
3635 ANYOF_CLASS_SET(data->start_class,ANYOF_DIGIT);
3637 for (value = 0; value < 256; value++)
3639 ANYOF_BITMAP_SET(data->start_class, value);
3644 if (flags & SCF_DO_STCLASS_AND) {
3645 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_DIGIT);
3646 for (value = 0; value < 256; value++)
3648 ANYOF_BITMAP_CLEAR(data->start_class, value);
3651 if (data->start_class->flags & ANYOF_LOCALE)
3652 ANYOF_CLASS_SET(data->start_class,ANYOF_NDIGIT);
3654 for (value = 0; value < 256; value++)
3655 if (!isDIGIT(value))
3656 ANYOF_BITMAP_SET(data->start_class, value);
3660 CASE_SYNST_FNC(VERTWS);
3661 CASE_SYNST_FNC(HORIZWS);
3664 if (flags & SCF_DO_STCLASS_OR)
3665 cl_and(data->start_class, and_withp);
3666 flags &= ~SCF_DO_STCLASS;
3669 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
3670 data->flags |= (OP(scan) == MEOL
3674 else if ( PL_regkind[OP(scan)] == BRANCHJ
3675 /* Lookbehind, or need to calculate parens/evals/stclass: */
3676 && (scan->flags || data || (flags & SCF_DO_STCLASS))
3677 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM)) {
3678 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
3679 || OP(scan) == UNLESSM )
3681 /* Negative Lookahead/lookbehind
3682 In this case we can't do fixed string optimisation.
3685 I32 deltanext, minnext, fake = 0;
3687 struct regnode_charclass_class intrnl;
3690 data_fake.flags = 0;
3692 data_fake.whilem_c = data->whilem_c;
3693 data_fake.last_closep = data->last_closep;
3696 data_fake.last_closep = &fake;
3697 data_fake.pos_delta = delta;
3698 if ( flags & SCF_DO_STCLASS && !scan->flags
3699 && OP(scan) == IFMATCH ) { /* Lookahead */
3700 cl_init(pRExC_state, &intrnl);
3701 data_fake.start_class = &intrnl;
3702 f |= SCF_DO_STCLASS_AND;
3704 if (flags & SCF_WHILEM_VISITED_POS)
3705 f |= SCF_WHILEM_VISITED_POS;
3706 next = regnext(scan);
3707 nscan = NEXTOPER(NEXTOPER(scan));
3708 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
3709 last, &data_fake, stopparen, recursed, NULL, f, depth+1);
3712 FAIL("Variable length lookbehind not implemented");
3714 else if (minnext > (I32)U8_MAX) {
3715 FAIL2("Lookbehind longer than %"UVuf" not implemented", (UV)U8_MAX);
3717 scan->flags = (U8)minnext;
3720 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3722 if (data_fake.flags & SF_HAS_EVAL)
3723 data->flags |= SF_HAS_EVAL;
3724 data->whilem_c = data_fake.whilem_c;
3726 if (f & SCF_DO_STCLASS_AND) {
3727 const int was = (data->start_class->flags & ANYOF_EOS);
3729 cl_and(data->start_class, &intrnl);
3731 data->start_class->flags |= ANYOF_EOS;
3734 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
3736 /* Positive Lookahead/lookbehind
3737 In this case we can do fixed string optimisation,
3738 but we must be careful about it. Note in the case of
3739 lookbehind the positions will be offset by the minimum
3740 length of the pattern, something we won't know about
3741 until after the recurse.
3743 I32 deltanext, fake = 0;
3745 struct regnode_charclass_class intrnl;
3747 /* We use SAVEFREEPV so that when the full compile
3748 is finished perl will clean up the allocated
3749 minlens when its all done. This was we don't
3750 have to worry about freeing them when we know
3751 they wont be used, which would be a pain.
3754 Newx( minnextp, 1, I32 );
3755 SAVEFREEPV(minnextp);
3758 StructCopy(data, &data_fake, scan_data_t);
3759 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
3762 SCAN_COMMIT(pRExC_state, &data_fake,minlenp);
3763 data_fake.last_found=newSVsv(data->last_found);
3767 data_fake.last_closep = &fake;
3768 data_fake.flags = 0;
3769 data_fake.pos_delta = delta;
3771 data_fake.flags |= SF_IS_INF;
3772 if ( flags & SCF_DO_STCLASS && !scan->flags
3773 && OP(scan) == IFMATCH ) { /* Lookahead */
3774 cl_init(pRExC_state, &intrnl);
3775 data_fake.start_class = &intrnl;
3776 f |= SCF_DO_STCLASS_AND;
3778 if (flags & SCF_WHILEM_VISITED_POS)
3779 f |= SCF_WHILEM_VISITED_POS;
3780 next = regnext(scan);
3781 nscan = NEXTOPER(NEXTOPER(scan));
3783 *minnextp = study_chunk(pRExC_state, &nscan, minnextp, &deltanext,
3784 last, &data_fake, stopparen, recursed, NULL, f,depth+1);
3787 FAIL("Variable length lookbehind not implemented");
3789 else if (*minnextp > (I32)U8_MAX) {
3790 FAIL2("Lookbehind longer than %"UVuf" not implemented", (UV)U8_MAX);
3792 scan->flags = (U8)*minnextp;
3797 if (f & SCF_DO_STCLASS_AND) {
3798 const int was = (data->start_class->flags & ANYOF_EOS);
3800 cl_and(data->start_class, &intrnl);
3802 data->start_class->flags |= ANYOF_EOS;
3805 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3807 if (data_fake.flags & SF_HAS_EVAL)
3808 data->flags |= SF_HAS_EVAL;
3809 data->whilem_c = data_fake.whilem_c;
3810 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
3811 if (RExC_rx->minlen<*minnextp)
https://perl5.git.perl.org / perl5.git / blob